Journal of Emerging Technologies in Web Intelligence
ISSN 1798-0461
Volume 4, Number 2, May 2012
Contents
Special Issue: Selected Best Papers of the International Conference on Information and
Communication Systems (ICICS 09)
Guest Editors: Mohammad AL-Rousan
Guest Editorial
Mohammad AL-Rousan
117
SPECIAL ISSUE PAPERS
The Boundaries of Natural Language Processing Techniques in Extracting Knowledge from Emails
Thomas W. Jackson, Sara Tedmori, Chris J. Hinde, and Anoud I. Bani-Hani
119
MALURLS: A Lightweight Malicious Website Classification Based on URL Features
Monther Aldwairi and Rami Alsalman
128
Vision-based Presentation Modeling of Web Applications: A Reverse Engineering Approach
Natheer Khasawneh, Oduy Samarah, Safwan Al-Omari, and Stefan Conrad
134
Adaptive FEC Technique for Multimedia Applications Over the Internet
Mohammad AL-Rousan and Ahamd Nawasrah
142
Semantic Approach for the Spatial Adaptation of Multimedia Documents
Azze-Eddine Maredj and Nourredine Tonkin
148
Adaptive Backoff Algorithm for Wireless Internet
Muneer O. Bani Yassein, Saher S. Manaseer, and Ahmad A. Momani
155
ConTest: A GUI-Based Tool to Manage Internet-Scale Experiments over PlanetLab
Basheer Al-Duwairi, Mohammad Marei, Malek Ireksossi, and Belal Rehani
164
Context-oriented Software Development
Basel Magableh and Stephen Barrett
172
Aggregated Search in XML Documents
Fatma Zohra Bessai-Mechmache and Zaia Alimazighi
181
The Developing Economies’ Banks Branches Operational Strategy in the Era of E-Banking: The Case
of Jordan
Yazan Khalid Abed-Allah Migdadi
189
Evolving Polynomials of the Inputs for Decision Tree Building
Chris J. Hinde, Anoud I. Bani-Hani, Thomas.W. Jackson, and Yen P. Cheung
198
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
117
Special Issue on Selected Best Papers of the International Conference on Information
and Communication Systems (ICICS 09)
Guest Editorial
The International Conference on Information and Communications Systems (ICICS2011) is a forum for Industry
Professionals and Academics from Companies, Governmental Agencies, and Universities around the world to present
their latest research results, ideas, developments and applications in all areas of Computer and Information Sciences.
The topics that have been covered in the ICICS2011 include, but are not limited to: Artificial Intelligence, Mobile
Computing, Networking, Information Security and Cryptography, Web Content Mining, Bioinformatics and IT
Applications, Database Technology, Systems Integration, Information Systems Analysis and Specification, and
Telecommunications. We selected 11 high quality papers (out of 54 papers, which were presented at the ICICS2-11)
and invited the authors of the selected papers to extend them and submit them for a complete new peer-review for
consideration in this Special Issue (SI). The final decision for the inclusion in the SI has been strictly based on the
outcome of the review process. The main objective of the SI is to make available the latest results in the field to the
research community and report state-of-the-art and in-progress research on all aspects of information and
communication systems. The selected papers span a broad range on the information retrieval, E-business and Internet.
The contributions of these papers are outlined below.
Jackson et. al, have studied the boundaries of natural language processing techniques in extracting Knowledge from
emails, where they aimed to determine if natural language processing techniques can be used to fully automate the
extraction of knowledge from emails. Based on the system built by the authors and it has been shown that although the
f-measure results are world leading, there is still a requirement for user intervention to enable the system to be accurate
enough to be of use to an organisation. On the hand, Al-Dwairi and Alsalman fcused on a very major problem in the
World Wide Web where they proposed a lightweight system to detect malicious websites online based on URL lexical
and host features and call it MALURLs. The system relies on Naïve Bayes classifier as a probabilistic model to detect if
the target website is a malicious or benign. It introduces new features and employs self learning using Genetic
Algorithm to improve the classification speed and precision. The system achieves an average precision of 87%.
Two interesting studies have been presented for Web applications; The first one is vision-based presentation
modelling of web applications, by Khasawneh et. al. This valuable work discussed the design, implementation, and
evaluation of a reverse engineering tool that extracts and builds appropriate UML presentation model of existing Web
applications. Their approach relies on a number of structured techniques such as page segmentation, the approach was
applied and evaluated in the Goalzz home page. The Second work is presented by AL-Rousan and Nawasrah on
techniques for Forward Error Correction (FEC) multimedia streams over the Internet. The work proposed a new
approach of adaptive FEC scheme that optimized the redundancy of the generated codewords from a Reed-Solomon
(RS) encoder. The adaptation of the FEC scheme is based on predefined probability equations, which are derived from
the data loss rates related to the recovery rates at the clients. Along the same line, Maredj and Tonkin proposed a new
generic approach to the spatial adaptation of multimedia documents. This approach allows heterogeneous devices
(desktop computers, personal digital assistants, phones, via Internet, etc.), to play multimedia documents under various
constraints (small screens, low bandwidth). This approach will enhance the access of multimedia content over the
Internet regardless of their specifications
Next, an interesting backoff algorithm for wireless Internet is proposed by Bani Yassein et al. Unlike the Binary
Exponential Backoff algorithm which makes exponential increments to contention window sizes, their work studied the
effect of choosing a combination between linear, exponential and logarithmic increments to contention windows.
Results have shown that choosing the right increment based on network status enhances the data delivery ratio up to
37% compared to the Binary Exponential Backoff, and up to 39 % compared to the Pessimistic Linear Exponential
Backoff algorithms for wireless Internet. A study presented by Al-Duwairi et al., introduced a new GUI-based tool to
manage Internet-Scale experiments over PlanetLab PlanetLab is being used extensively to conduct experiments,
implement and study large number of applications and protocols in Internet-like environment. The tool (called ConTest)
enables PlanetLab users to setup experiment and collect results in a transparent and easy way. This tool also allows
different measurements for different variables over the PlanetLab network.
Content-oriented programming, searching, and retrieval are emerging techniques that enables these systems be more
context aware. Magableh and Barrett described a development paradigm for building context-oriented applications
using a combination of Model-Driven Architecture that generates an ADL, which presents the architecture as a
components-based system, and a runtime infrastructure (middleware) that enables transparent self-adaptation with the
underlying context environment. On the other hand, Zohra et. al. proposed an aggregated approach for searching a
content of an XML documents. The objective of their work is to build a virtual elements that contain relevant and nonredundant elements, that are likely to answer better the query that elements taken separately. In related issue, e-banking
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.117-118
118
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
systems involves heavy retrieval and content-based transactions. Yazan Migdadi in his paper has addressed this issue
with a particular attention paid to the e-banking in Jordanian.
Finally, as guest –co-editors of this SI, we would like to express our deepest thanks to the Editor-in-Chief, Professor
Sabah Mohammed for hosting this Issue in the JETWI and for his continued support and helpful guidance throughout
all the stages of preparing this SI. Our sincere thanks also go to the Editorial-office staff of the journal for their
excellent job during the course of preparing this special issue. We also thank the authors for their contributions,
including those whose papers were not included. We thank and greatly appreciate the thoughtful work of many
reviewers who provided invaluable evaluations and recommendations.
Guest Editor
Mohammad AL-Rousan
College of Computer and Information Technology, Jordan University of Science & Technology, Jordan
Email: alrousan@just.edu.jo
Mohammad Al-Rousan is currently A full professor at the Department of Network Engineering and
Security, Jordan University of Science and Technology (JUST). He was educated in KSA, Jordan and USA.
He received his BSc in Computer Engineering from King Saud University, Saudi Arabia, in 1986. He
received his M.S. in Electrical and Computer Engineering from University of Missouri-Columbia, MI, USA
in 1992. In 1996 , he was awarded the PhD in Electrical and Computer Engineering from Brigham Young
University, UT, USA. He was then an assistant professor at JUST, Jordan. In 2002, he joint the Computer
Engineering Department at American University of Sharjah, UAE. Since 2008, he has been the Dean of
College of Computer and Information Technology at JUST. He is the Director of the Artificial Inelegant and
Robotics Laboratory, and a co-founder for the Nano-bio laboratory, at JUST. His search interests include
wireless networking, System protocols, intelligent systems, computer applications, and Nanotechnology,
Internet computing. Dr. Al-Rousan served on organising and program committees for many prestigious
international conferences. He is the recipient of several prestigious awards and recognitions. He co-chaired international conferences
on Information and Communication Systems (ICICS09 and ICICS11).
© 2012 ACADEMY PUBLISHER
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
119
The Boundaries of Natural Language Processing
Techniques in Extracting Knowledge from
Emails
Thomas W. Jackson
Information Science, Loughborough University, UK
Email: T.W.Jackson@lboro.ac.uk
Sara Tedmori
Computer Science Department, Princess Sumaya University for Technology, Jordan
Email: S.Tedmori@psut.edu
Chris J. Hinde and Anoud I. Bani-Hani
Computer Science, Loughborough University, UK
Email: C.J.Hinde@lboro.ac.uk; A.I.Bani-Hani@lboro.ac.uk
Abstract—The aim of this research is to determine if natural
language processing techniques can be used to fully
automate the extraction of knowledge from emails. The
paper reviews the five generations of building systems to
share knowledge and highlights the challenges faced by all.
The paper discuss the system built by the authors and shows
that although the f-measure results are world leading, there
is still a requirement for user intervention to enable the
system to be accurate enough to be of use to an organisation.
Index Terms—Natural Language Processing, Keyphrase
Extraction, f-measure, Email Communication, Knowledge
Extraction, NLP Limitations
I. INTRODUCTION
Over the last several decades, many reports [1], [2],
[3], [4], [5] have indicated that people searching for
information prefer to consult other people, rather than to
use on-line or off-line manuals. Allen [5] found that
engineers and scientists were roughly five times more
likely to consult individuals rather than impersonal
sources such as a database or file cabinet for information.
In spite of the advancements in computing and
communications technology, this tendency still holds;
people remain the most valued and used source for
knowledge[6], [7].
Unfortunately, finding individuals with the required
expertise can be extremely expensive [8], [9], as it is time
consuming and can interrupt the work of multiple persons.
A common problem with many businesses today, large
and small, is the difficulty associated with identifying
where the knowledge lies. A lot of data and information
generated and knowledge gained from projects reside in
the minds of employees. Therefore the key problem is,
how do you discover who possesses the knowledge
sought?
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.119-127
In the search for the solution, information systems
have been identified as key players with regards to their
ability to connect people to people to enable them to
share their expertise and collaborate with each other [10],
[11], [12], [13]. Thus, the solution is not to attempt to
archive all employees’ knowledge, but to link questions
to answers or to knowledgeable people, who can help
find the answers sought [14]. This has led to the interest
in systems, which help connect people to others that can
help them solve their problems, answer their questions,
and work collaboratively.
Cross et al. [15] reviewed [16], [17], [18], [19], [20],
[21], [22], [23], [24], and summarises the benefits of
seeking information from other people. These benefits
include:
• provision of solutions to problems;
• provision of answers to questions;
• provision of pointers to others that might know
the answer
• provision of pointers to other useful sources;
• engagement in interaction that helps shape the
dimension of the problem space;
• psychological
benefits (e.g. confidence,
assurance);
• social benefits (e.g. social approval for decisions,
actions);
• improvement in the effectiveness with which a
person advances their knowledge in new and
often diverse social contexts;
• improvement in efficiency (e.g. reduction in
time wasted pursuing other avenues); and
• legitimation of decisions.
Cross [15] identifies five categories that these benefits
fall under: (1) solutions (know what and know how); (2)
meta-knowledge (pointers to databases or other people);
120
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
(3) problem reformulation; (4) validation of plans or
solutions; and (5) legitimation from contact with a
respected person. It has been recognised that the idea of
connecting people to people is a way forward, yet from a
natural language processing viewpoint what has been
attempted before and what are the limitations of the
current systems.
This paper reviews the expert finding approaches and
discusses the natural language processing (NPL)
techniques used to extract knowledge from email,
including the one developed by the authors. It concludes
by reflecting on the current f-measure scores for
knowledge extraction and the role of the user in any
knowledge location system.
II. EMAIL, KNOWLEDGE WORK,
AND EXPERTISE LOCATION
This section explores the role of email in knowledge
work, focusing on its potential value for expertise
location. Email is an important knowledge channel [47],
and collaboration tool [48], actively used by
organisational knowledge workers worldwide. Email has
been reported as the most common Internet activity, used
daily by ninety-one percent of U.S. internet users [49].
The organisational dominance of email is also
demonstrated by a recent study reporting that email
messages involve an estimated average of 15.8
Megabytes of archive storage per corporate end-user per
day[50].
Despite email’s popularity and ubiquity, there is little
research on the value that email provides to
organisational knowledge management (KM) strategies.
Email enables greater knowledge work than possible in
earlier technological eras [51], [52]. It enables knowledge,
creation [53], [54], knowledge sharing and knowledge
flow [55]. According to Tedmori et al., employees are
motivated to use email for knowledge work for reasons
including [56]:
•
•
•
•
•
•
•
•
Email messages attract workers’ attention;
Email is well integrated with everyday work;
Email discourse provides a context for sensemaking about ideas, projects and other types of
business knowledge;
Email enables the referencing of work objects
(such as digital documents), and provides a
history via quoted messages;
Email’s personalized messages are appealing,
meaningful and easily understood;
Email
encourages
commitment
and
accountability by automatically documenting
email exchanges;
Email is collected in inboxes and organisational
archives, email represents valuable individual,
collective and organisational memories that may
be tapped later; and
Email discourse facilitates the resolution of
multiple conflicting perspectives which can
© 2012 ACADEMY PUBLISHER
stimulate an idea for a new or improved process,
product or service.
Email provides several important, often unexploited,
opportunities for expertise-finding. Knowledge in email
can be accessed and reused directly [57] or can serve
indirectly as a pointer to an expert [30], [58], [56]. A
recognized definition of an expert is someone who
possesses specialised skills and knowledge derived from
training and experience [59].
Traditionally, email clients were designed for the
reuse of personal knowledge archives. For example,
folders are popular structures for organising email
messages so that they assist owners with knowledge reuse.
This facility was highlighted by a recent study of Enron’s
publicly available email archive, where significant folder
usage was employed [60]. Employees often search
personal email archives seeking knowledge, in preference
to searching electronic knowledge repositories (EKR)
[57], raising questions about the effectiveness of EKRs
for reuse, an issue first raised by [61]. As mentioned
earlier, Swaak et al.’s study also found that employees
prefer to find an expert to help them with their
knowledge-based concern, rather than searching directly
for knowledge [57]. The next section describes automated
expert finder tools that exploit email as evidence of
expertise.
III. EXPERT FINDING APPROACHES
Various approaches to expertise location have been
developed and implemented to link expertise seekers with
internal experts. These have been initially discussed in
Lichtenstein et al.’s research [62], but in this paper the
authors extend the review to include the fifth generation
of extraction systems.
The first generation of such systems sprung out of the
use of helpdesks as formal sources of knowledge, and
comprised knowledge directories and expert databases.
Microsoft’s SPUD project, Hewlett-Packard’s CONNEX
KM system, and the SAGE expert finder are key
examples of this genre. Generally expert databases have
‘Yellow Pages’ interfaces representing electronic
directories of experts linked to their main areas of
expertise. Such directories are based on expert profiles
which must be maintained by experts on a voluntary basis.
The key advantages of such directories include
conveniently connecting those employees inadequately
tapped into social and knowledge networks with relevant
experts. However such approaches also suffer from
significant shortcomings. Unless employees regularly
update their profiles, the profiles lose accuracy and no
longer reflect reality. Yet employees are notorious for
neglecting to update such profiles as such duties are often
considered onerous and low priority [25]. Employees
may not wish to provide expertise. Overall, when large
numbers of employees are registered and profiles are
inaccurate, credibility is rapidly lost in such systems
which are increasingly ignored by knowledge seekers,
who instead rely on social networks or other methods [9].
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
In addition, expertise descriptions are usually incomplete
and general, in contrast with the expert-related queries
that are usually fine-grained and specific, and replete with
various qualitative requirements [25].
In the second generation of expertise locators,
companies took advantage of personal web pages where
employees could advertise expertise internally or
externally. Such pages are designed according to
corporate templates or personal design, and are usually
made accessible via the World Wide Web or corporate
intranets. The convenience of web site creation and
update, web site retrieval and access, and sophisticated
search engines, are key advantages of this approach.
However, employees may lack the motivation, time or
technical expertise to develop or update their profiles,
which rapidly lose accuracy and credibility and the
capacity to meet expert location needs [25]. In addition,
as noted by Yimam-Seid and Kobsa [25], employee use
of search engines for locating an expert’s web page may
be ineffective since such a process is based on a simple
keyword matching task which does not always yield the
most relevant experts’ web pages. The search activity can
also be very time consuming when a high number of hits
is returned and an employee must then systematically or
randomly attempt to choose and explore the listed link(s).
As Yimam-Seid and Kobsa have observed for this
approach, knowledge seekers are allocated significant and
often onerous responsibility for finding relevant experts
([25]. The second generation of approaches also included
the development of more dynamic expert databases.
Answer Garden [26], [27], which is a question-answering
system, maintains a database of frequently asked
questions and answers. When the system does not find
required information in the database, an end-user may ask
the question of the system. Answer Garden then routes
the question to the corresponding experts. However, it is
not clear with this approach how the system identifies
experts and, in particular, whether experts have
nominated their own areas and levels of expertise.
The third generation of approaches relies primarily on
secondary sources for expert identification. For example,
the web application Expertise Browser [28], studies
browsing patterns/activities in order to identify experts.
With this application, if the user knows a particular
expert, the user can ask the system to reveal the browsing
path of that expert, relevant to the user’s query. Among
other disadvantages, if an employee does not know an
expert, the user must ask the system to identify one or
more experts. The employee must then scan the browsing
paths of the identified experts for possibly useful links,
which can be a very time consuming process.
Furthermore, it is likely that browsing reveals interests
rather than expertise [25]. The monitoring of browsing
patterns clearly involves privacy issues that such systems
fail to address. Other secondary-source approaches utilise
message board discussions as indicators of expertise. For
example, ContactFinder [29] is a research prototype that
reviews messages posted on message boards.
ContactFinder analyses subject areas from messages and
links them to the names of experts who wrote the
© 2012 ACADEMY PUBLISHER
121
messages. It provides users seeking experts with expert
referrals when user questions match expert’s earlier
postings. All such approaches infer experts from
secondary sources but do not allow experts to confirm
such inferences.
A recently recognised socially based approach is the
use of social networks which provide a complex social
structure for the development of social capital and the
connection of novices and experts [11]. In a study
conducted by [7], while some people were difficult to
access, they were still judged to be valuable sources of
help. The use of a social network to locate expertise has
become popular because colleagues are often physically
available, are personal friends, or are known to be experts
on the topic. However, there is no guarantee that a
genuine expert will be consulted, as users may choose to
consult a moderately knowledgeable person, a person
with whom a good relationship exists, a proximate
employee, or a quickly located employee, simply as that
person is within the expertise seeker’s social network.
With this approach, low quality expertise may be
systematically input into an organisation where it is
quickly applied. Automated social network approaches
such as Referral Web suffer from similar concerns.
The fourth generation may include one or more the
above approaches together with natural language
processing and artificial intelligence techniques in order
to analyse stored knowledge, seeking to identify expertise
and experts [25], [30], [31]. A forerunner of such
systems was Expert Locator which returns pointers to
research groups in response to natural language queries
on reports and web pages [32]. A second example is
Expert Finder [33] which considers self-published
documents containing the topic keyword, and the
frequency of the person named near the same topic
keyword in non-self-published documents, in order to
produce expertise scores and ranks. In 1993 Schwartz and
Wood first attempted to utilise e-mail messages, known
to be heavily knowledge-based, to deduce shared-interest
relationships between employees. In 2001, following
other experts’ promising attempts, Sihn & Heeren
implemented XpertFinder, the first substantial attempt to
exploit the knowledge-based content of e-mail messages
by employing technology to analyse message content.
More recently Google Mail have use similar techniques
to scan email content whilst reading messages on-line, to
extract key phrases that can then be matched with specific
marketing adverts that appear to the right hand side of the
browser. This is more a case of just-in-time knowledge
that could be extremely useful to employees if, for
example, they were writing reports and the application
would mine for keywords and link the user to existing
written material or experts to aid in the report writing task.
The drawback of many of the fourth generation
approaches is the output. For example, expert listings are
unordered when presented to a user seeking experts,
which means significant user effort is required to identify
the best expert. Such systems identify experts by textual
analysis but rarely support expert selection by users. In
addition, such systems fail to present the varying degrees
122
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
(or levels) of expertise that people possess and tend to
assume a single level of expertise. It is thus entirely the
user’s responsibility to systematically process the
returned results in order to identify the most suitable
experts for answering specific queries. Techniques
employed to build the fourth generation expertise profiles
should be advanced to ensure that the textual fragments
analysed accurately convey employees’ expertise. To date,
the automated techniques have been inadequate because
they cannot distinguish between what is important and
what is not important in identifying an expert. In addition,
the system should be able to match user needs with
expertise profiles by using appropriate retrieval
techniques, ensuring that relevant experts are not
overlooked and that less relevant experts are not
overburdened with inappropriate queries.
Numerous attempts have been made by researchers in
both academia and industry to semi-automate/ automate
the process of finding the right expert. What could be
deemed as the fifth generation of systems, work by
analyzing who we socialize with. IBM’s SmallBlue
project (which is also part of the Atlas software suite)
applies artificial intelligence algorithms for purpose of
uncovering participants’ social network “who they know”
and also the expertise of those participants “what they
know” [63]. The application analyses employees’ email
outboxes, but only for the employees that work at the
same company. The application also analyses employees’
outgoing instant messaging chat transcripts and their
profile information. The inferred social network for each
participant is private and displayed as a visualization in
Ego for each user. SmallBlue comes with a search engine
that enables users to search for expertise based on a topic
and displays the results in the context of the wider
enterprise social network. There are many other systems
available that could be deemed part of the fifth generation
of systems, e.g. AskMe, Autonomy's IDOL, Triviumsoft's
SEE-K, MITRE’s ExpertFinder, MITRE’s XpertNet,
MindServer Expertise, but they also have similar
limitations to those that have been discussed earlier in
this section.
This extended but abbreviated evolutionary review of
expertise locator systems has highlighted the need for
new expert locator systems with enhanced information
retrieval techniques that provide user friendly expertise
seeking techniques and high levels of accuracy in
identifying relevant experts. In the next section, we
summarise the techniques that have been used to extract
key phrases and then discuss the latest attempts by the
authors to improve upon the techniques to enhance the
accuracy of the key phrases extracted and the ranking of
their importance according to a user’s expertise.
IV. KEY PHRASE EXTRACTION
Numerous papers explore the task of producing a
document summary by extracting key sentences from the
document [34], [35], [36], [37], [38], [39], [40], [41], [42],
[43].
© 2012 ACADEMY PUBLISHER
The two main techniques are domain dependent and
domain independent. Domain dependent techniques
employ machine learning and require a collection of
documents with key phrases already attached, for training
purposes. Furthermore, the techniques (both domain
dependent and domain independent) are related to
linguistics and/or use pure statistical methods. A number
of applications have been developed using such
techniques. A full discussion of existing approaches,
together with their merits and pitfalls, is provided in [44].
There are many weaknesses with current approaches
to automatic key phrase identification, several of which
are discussed here to illustrate the issues. First, the
extraction of noun phrases from a passage of text is
common to all such approaches [43], [45]. However, a
disadvantage of the noun extraction approach is that,
despite the application of filters, many extracted key
phrases are common words likely to occur in numerous emails in many contexts. Therefore, it is important to
distinguish between more general nouns and nouns more
likely to comprise of key phrases. Second, Hulth [45]
pinpoints two common drawbacks with existing
algorithms, such as KEA. The first drawback is that the
number of words in a key phrase is limited to three. The
second drawback is that the user must state the number of
keywords to extract from each document [45].
In the attempt to push the boundaries of key phrase
extraction, work undertaken by the authors aimed to
enable end-users to locate employees who may possess
specialised knowledge that users seek. The underlying
technical challenge in utilising e-mail message content
for expert identification is the extraction of key phrases
indicative of sender skills and experience.
In developing the new system the Natural Language
ToolKit (NLTK) was employed to build a key phrase
extraction “engine”. NLTK comprises a suite of libraries
and programs for symbolic and statistical natural
language processing (NLP) for the Python programming
language. The completed key phrase extractor was then
embedded within EKE - an Email Knowledge Extraction
process based on two stages.
The first stage involves a training process which
enables the creation of a speech-tagging model for
tagging parts-of-speech (POS) within an e-mail message.
The second stage involves the extraction of key phrases
from e-mail messages with the help of the speech-tagging
model. Figure 1 shows the keyphrase extraction steps of
stage two. The email message sent by the user is captured
by the system and the text in the email body is fed into
the keyphrase extraction engine. The text is then divided
into tokens using regular expression rules and tagged by
their parts of speech with the help of the POS model
created. Keyphrases are identified by applying rules
which were manually set by the author to group all
occurrences of specific sequences of tags together. The
rule is formed from a sequence of grammatical tags
which are most likely to contain words that make up a
keyphrase. Once sequences of tags are collated, more
rules are applied to remove a subset of non relevant
phrases. Keyphrases are then chosen from the identified
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
candidate phrases. The process concludes with the use of
linguistic filtering to extract the most important
keyphrases. This results in a set of lines, each containing
a sequence of tokens representing key competencies.
Figure 2 depicts how the EKE system analyses e-mail
messages to identify experts (but many other systems use
a similar process). Once a message is sent by a user (step
1), the body of the message is captured by EKE.
Figure 2 Stages of the extraction process
EKE’s key phrase extraction engine will parse the
body of the email seeking appropriate key phrases that
might represent the user’s expertise (step 2). This process
is fully automated and takes only milliseconds to
complete, and is so far transparent to both sender and
receiver. It is possible that key phrases will not be
identified by the key phrase extraction engine as the
message may not contain any text suggesting key phrases,
or the message contains key phrases that were not
detected. In such cases, EKE will not require any action
from the user whose work activities will therefore remain
uninterrupted.
In step 3, if the engine identifies key phrases the user
is requested to rank the extracted key phrase using a scale
of 1 - 4, to denote level of user expertise in the
corresponding field. The rankings 1 – 4 represent basic
knowledge, working knowledge, expert knowledge, or
not applicable. The four point categorisation scale was
devised because a seeker of knowledge should be
forewarned that a self-nominated expert may lack an
expected capability. The knowledge seeker can then
decide whether to proceed to contact such an expert for
help. In Figure 1, “Questionnaire”, “Semantics”, “Casino”
and “Online database” are examples of the key phases
that have been extracted from the body of a message. On
© 2012 ACADEMY PUBLISHER
123
average very few key phrases are extracted from a
message because generally, according to our development
tests and pilot studies, there are few key-phrases
contained within any one e-mail message. Therefore
typically a user is not unduly delayed by the key phrase
expertise categorisation process. Once categorised (for
example in Figure 1, “Questionnaire” may be categorised
as basic knowledge, “Semantics” as expert knowledge,
and so on), key phrases are stored in an expertise profile
database (excluding key phrases categorised as “not
applicable”). The user can access and edit his/her expert
profile at anytime (step 4). The key phrases that are
stored in the expertise profile database are also made
available to other employees within the organisation, to
enable them to locate relevant experts by querying the
database (step 5).
Figure 2 – Overview of the E-mail Knowledge
Extraction System
The EKE system has significant advantages compared
with other e-mail key phrase extraction systems, not all of
which perform steps 3 and 4. The present system gains
accuracy by requiring a user in steps 3 and 4 to rank his
or her level of expertise for a particular key phrase. Most
existing systems attempt to rank experts automatically
rather than consulting users for their perceptions of their
level of expertise. Such systems are likely to be less
successful at accurately identifying expertise levels as
they do not capture employee knowledge of their own
expertise. The above approach has been trialled at
Loughborough University and has shown to be effective
in correctly identifying experts [44]. However,
it is
important to note that this system used a hybrid approach
of NLP and user intervention to determine the usefulness
of the key phrases extracted. The focus of this paper is
reviewing the boundaries of NLP techniques and the next
section reviews the results of the NLP system without
user intervention, which leads to a discussion about the
boundaries of NLP in key phrase extraction.
V. RESULTS AND BOUNDARIES OF NLP
It is important to note that the full EKE system which
includes user intervention has not been included in these
124
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
results. The full results have been published in Tedmori’s
et al.’s research [44]. The purpose of showing the results
in this paper is to illustrate that the f-measure results have
not improved over a number of years. This highlights that
the discipline has potential found the boundaries of NLP
in extracting knowledge keyphrases from email. The
Natural Language ToolKit system developed by the
authors was tested on a number of corpuses.
•
•
•
Corpus 1 - Emails from various academic
domains; Size 45
Corpus 2 - Total office solutions organisation;
Size 19
Corpus 3 – Enron; Size 50
The sampling units were collected from subjects from
different backgrounds (people with English as their first
language and people who can communicate in English,
but is not their first language). All subjects belong to the
age group 24-60. All the sampling units were outgoing
mail. The authors believe that sampling units are
representative of typical messages that are sent out in
institutional and corporate environments. The sampling
units of the sample, Corpus 1, were collated from various
academic disciplines (computer science, information
science, building and construction engineering). The
sampling units of the second sample, Corpus 2, are
specific to one employee from a large supplier of total
office solutions in the UK & Ireland, which for
confidentiality reasons in is referred to as Employee E
from Company XYZ. The sampling units of the final
sample, Corpus 3, are collated from the Enron email
dataset, which is freely available on the net.
The f-measure, a widely used performance measure in
information retrieval, was used to measure the system and
is defined by (1).
Equation 1 – F-Measure Calculation
where precision is the estimate of the probability that if a
given system outputs a phrase as a key phrase, then it is
truly a key phrase and recall is an estimate of the
probability that, if a given phrase is a key phrase, then a
given system will output it as a key phrase.
Table I – Results of testing the author’s Natural Language ToolKit
system
Corpus
Precision
Recall
Corpus 1
Corpus 2
Corpus 3
53.3
59.6
41.7
57.6
63.1
48.3
fmeasure
55.4
61.3
44.8
In Table I, precision, recall, and the f-measure results
are shown. The highest precision (59.6), recall (63.1),
and f-measure (61.3) were achieved on the smallest
© 2012 ACADEMY PUBLISHER
sample (19 messages). Since only three sets were
evaluated, one cannot determine the coloration between
size of the sample and performance of the extractor.
Turney [47] evaluates four key phrase extraction
algorithms using 311 email messages collected from 6
employees, and in which 75% of each employee’s
messages was used for training and 25% (approximately
78 messages) was used for testing. His evaluation
approach is similar to the authors of this paper and the
highest f-measure reported was that of the NRC, the
extractor component of GenEx, which uses supervised
learning from examples. The f-measure reported is 22.5,
which is, as expected, significantly less than the fmeasures shown in Table I. Hulth [45] reports results
from three different term selection approaches. The
highest f-measure reported was 33.9 from the n-gram
approach with POS tags assigned to the terms as features.
All unigrams, bigrams, and trigrams were extracted, after
which a stop list was used where all terms beginning or
ending with a stopword were removed.
The Natural Language ToolKit system developed by
the authors appears to have the best f-measure results in
the world when it comes to email knowledge extraction.
Although the results are pleasing, the sight of a fully
automated system that can extract knowledge from email
without user intervention appears to be many years away,
if at all possible. However, with the financial muscle of
organizations like Google developing techniques for their
range of information retrieval applications, this domain is
likely to see rapid progress within a short period of time.
VI. CONCLUSION
This paper has reviewed the four generations of
building systems to share knowledge and highlighted the
challenges faced by all. The paper discussed the
techniques used to extract key phrases and the limitations
in the NLP approaches which have defined the
boundaries of the domain. The paper has shown that
although the f-measure results of the study are
encouraging, there is still a requirement for user
intervention to enable the system to be accurate enough to
provide substantial results to the end users. It is
concluded that NLP techniques are still many years away
from providing a fully automated knowledge extraction
system.
REFERENCES
[1]
[2]
[3]
[4]
[5]
Hiltz, S.R. (1985). Online Communities: A Case Study
of the Office of the Future, Ablex Publishing Corp,
Norwood, NJ.
Lang, K.N., Auld, R. & Lang, T. (1982). "The Goals
and Methods of Computer Users", International Journal
of Man-Machine Studies, vol. 17, no. 4, pp. 375-399.
Mintzberg, H. (1973). The Nature of Managerial Work,
Harper & Row, New York.
Pelz, D.C. & Andrews, F.M. (1966). Scientists in
Organizations: Productive Climates for Research and
Development, Wiley, New York.
Allen, T. (1977). Managing the Flow of Technology,
MIT Press, Cambridge, MA.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
Cross, R. & Sproull, L. (2004). "More Than an Answer:
Information Relationships for Actionable Knowledge",
Organization Science, vol. 15, no. 4, pp. 446-462.
Kraut, R.E. & Streeter, L.A. (1995). "Coordination in
Software Development", Communications of the ACM,
vol. 38, no. 3, pp. 69-81.
Maltzahn, C. (1995). "Community Help: Discovering
Tools and Locating Experts in a Dynamic Environment",
CHI '95: Conference Companion on Human Factors in
Computing SystemsACM, New York, NY, USA, pp.
260.
Campbell, C.S., Maglio, P.P., Cozzi, A. & Dom, B.
(2003). "Expertise Identification Using Email
Communications", Twelfth International Conference on
Information and Knowledge Management New Orleans,
LA, pp. 528.
Bishop, K. (2000). "Heads or Tales: Can Tacit
Knowledge Really be Managed", Proceeding of ALIA
Biennial Conference Canberra, pp. 23.
Cross, R. & Baird, L. (2000). "Technology is not
Enough: Improving Performance by Building
Organizational Memory", Sloan Management Review,
vol. 41, no. 3, pp. 41-54.
Gibson, R. (1997). Rethinking the Future: Rethinking
Business, Principles, Competition, Control &
Complexity, Leadership, Markets, and the World,
Nicholas Brealey, London.
Lang, J.C. (2001). "Managing in Knowledge-based
Competition", Journal of Organizational Change
Management, vol. 14, no. 6, pp. 539-553.
Stewart, T.A. (1997). Intellectual Capital: The New
Wealth of Organizations, Doubleday, New York, NY,
USA.
Cross, R. (2000). "More than an Answer: How Seeking
Information Through People Facilitates Knowledge
Creation and Use", Toronto, Canada.
Burt, R.S. (1992). Structural Holes: The Social
Structure of Competition. Harvard University Press,
Cambridge.
Erickson, B.H. (1988). "The Relational Basis of
Attitudes." in Social Structures: A Network Approach:,
Barry Wellman and S. D. Berkowitz (eds.), edn,
Cambridge University Press., New York:, pp. 99-121.
Schön, D.A. (1993). "Generative Metaphor: A
Perspective on Problem-setting in Social Policy" in
Metaphor and Thought, ed. A. Ortony, 2nd edn,
Cambridge University Press, Cambridge, pp. 137-163.
Walsh, J.P. (1995). "Managerial and Organizational
Cognition: Notes from a Trip down Memory Lane.",
Organizational Science, vol. 6, no. 3, pp. 280-321.
Weick, K.E. (1979). The Social Psychology of
Organising, 2nd edn, McGraw-Hill, New York.
Weick, K.E. (1995). Sense making in Organisations,
Sage, London.
Blau, P.M. (1986). Exchange and Power in Social Life,
Transaction Publishers, New Brunswick, NJ.
March, J.G. & Simon, H.A. (1958). Organizations,
Wiley, New York.
Lave, J. & Wenger, E. (1991). Situated Learning :
Legitimate Peripheral Participation, Cambridge
University Press, U.K.
Yimam-Seid, D. and Kobsa, A. (2003) ‘Expert finding
systems for organizations: problem and domain analysis
and the DEMOIR approach’, Journal of Organizational
Computing and Electronic Commerce, Vol. 13, No. 1,
pp.1–24.
© 2012 ACADEMY PUBLISHER
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
125
Ackerman, M.S. and Malone, T.W. (1990) ‘Answer
garden: a tool for growing organizational memory’,
Proceedings of ACM Conference on Office Information
Systems, Cambridge, Massachusetts, pp.31–39.
Ackerman, M.S. (1994) ‘Augmenting the organizational
memory: a field study of answer garden’, Proceedings
of the ACM Conference on Computer-Supported
Cooperative Work, pp.243–252.
Cohen, A.L., Maglio, P.P. and Barrett, R. (1998) ‘The
expertise browser: how to leverage distributed
organizational knowledge’, Presented at Workshop on
Collaborative Information Seeking at CSCW’98, Seattle,
Washington.
Krulwich, B. and Burkey, C. (1996a) ‘Learning user
information interests through the extraction of
semantically significant phrases’, In AAAI 1996 Spring
Symposium on Machine Learning in Information
Access, Stanford, California.
Balog, K. and de Rijke, M. (2007) ‘Determining expert
profiles (with an application to expert finding)’,
Proceedings of the Twentieth International Joint
Conferences on Artificial Intelligence, Hyderabad, India,
pp.2657–2662.
Maybury, M., D’Amore, R. and House, D. (2002)
‘Awareness of organizational expertise’, International
Journal of Human-Computer Interaction, Vol. 14, No. 2,
pp.199–217.
Streeter, L.A. and Lochbaum, K.E. (1988) ‘An
expert/expert-locating system based on automatic
representation of semantic structure’, Proceedings of the
Fourth
Conference
on
Artificial
Intelligence
Applications, San Diego, California, pp.345–349.
Mattox, D., Maybury, M. and Morey, D. (1999)
‘Enterprise expert and knowledge discovery’,
Proceedings of the 8th International Conference on
Human-Computer Interaction, Munich, Germany,
pp.303–307.
Luhn HP. 1958. The automatic creation of literature
abstracts. I.B.M. Journal of Research and Development,
2 (2), 159-165.
Marsh E, Hamburger H, Grishman R. 1984. A
production rule system for message summarization. In
AAAI-84, Proceedings of the American Association for
Artificial Intelligence, pp. 243-246. Cambridge, MA:
AAAI Press/MIT Press.
Paice CD. 1990. Constructing literature abstracts by
computer: Techniques and prospects. Information
Processing and Management, 26 (1), 171-186.
Paice CD, Jones PA. 1993. The identification of
important concepts in highly structured technical papers.
SIGIR-93: Proceedings of the 16th Annual International
ACM SIGIR Conference on Research and Development
in Information Retrieval, pp. 69-78, New York: ACM.
Johnson FC, Paice CD, Black WJ, Neal AP. 1993. The
application of linguistic processing to automatic abstract
generation. Journal of Document and Text Management
1, 215-241.
Salton G, Allan J, Buckley C, Singhal A. 1994.
Automatic
analysis,
theme
generation,
and
summarization of machine-readable texts. Science, 264,
1421-1426.
Kupiec J, Pedersen J, Chen F. 1995. A trainable
document summarizer. Proceedings of the 18th Annual
International ACM SIGIR Conference on Research and
Development in Information Retrieval, pp. 68-73, New
York: ACM.
126
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Brandow R, Mitze K, Rau LR. 1995. The automatic
condensation of electronic publications by sentence
selection. Information Processing and Management, 31
(5), 675-685.
Jang DH, Myaeng SH. 1997. Development of a
document summarization system for effective
information services. RIAO 97 Conference Proceedings:
Computer-Assisted Information Searching on Internet;
101-111. Montreal, Canada.
Tzoukermann E, Muresan S, Klavans JL. 2001. GISTIT: Summarizing Email using Linguistic Knowledge
and Machine Learning. In Proceeding of the HLT and
KM Workshop, EACL/ACL.
Tedmori, S., Jackson, T.W. and Bouchlaghem, D. (2006)
‘Locating knowledge sources through keyphrase
extraction’, Knowledge and Process Management, Vol.
13, No. 2, pp.100–107.
Hulth A. 2003. Improved Automatic Keyword
Extraction Given More Linguistic Knowledge. In:
Proceedings of the Conference on Empirical Methods in
Natural Language Processing (EMNLP'03). Sapporo.
Turney PD. 1997. Extraction of Keyphrases from
Text: Evaluation of Four Algorithms, National
Research Council, Institute for Information
Technology, Technical Report ERB-1051. (NRC
#41550)
Lichtenstein, S. 2004, "Knowledge Development
and Creation in Email", The 37th Annual Hawaii
International Conference on System Sciences
(HICSS'04) - Track 8IEEE Computer Society,
Washington, DC, USA.
Garcia, I. 2006, 04/03-last update, The Good in Email (or why Email is Still the Most adopted
Collaboration Tool) [Homepage of Central
Desktop],
[Online].
Available:
http://blog.centraldesktop.com/?entry=entry060403214628 [2011, 06/22].
Pew Internet & American Life Project 2007, 01/11last update, Pew Internet & American Life Project
Surveys, March 2000-April 2006 [Homepage of
Pew Research Center], [Online]. Available:
http://www.pewinternet.org/trends/Internet_Activiti
es_1.11.07.htm [2007, 11/27] .
Derrington, S. 2006, 11/11-last update, Getting
Control of the Storage Environment, [Online].
Available:
http://www.datastorageconnection.com/article.mvc/
Gaining-Control-Of-The-Storage-Environment-iB0002 [2011, 06/22] .
Jackson, T. & Burgess, A. 2003, "Capturing and
managing email knowledge.", Business Innovation
in the Knowledge Economy - Abstracts from the
IBM \& Stratford-Upon-Avon Conference, eds. J.
Abbot, L. Martin, R. Palmer, M. Stone & L.T.
Wright, , pp. 28.
Whittaker, S., Bellotti, V. & Moody, P. 2005,
"Introduction to this special issue on revisiting and
reinventing e-mail", Human-Computer Interaction,
vol. 20, pp. 1-9.
Ducheneaut, N. & Belloti, V. 2003, "Ceci n’est pas
un objet? Talking about objects in e-mail", HumanComputer Interaction, vol. 18, pp. 85-110.
Jackson, T.W. & Tedmori, S. 2004, "Capturing and
Managing Electronic Knowledge: The Development
of the Email Knowledge Extraction", The
International Resource Management Association
ConferenceIdea Group, New Orleans, USA, pp. 463.
© 2012 ACADEMY PUBLISHER
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
Bontis, N., Fearon, M. & Hishon, M. (2003). "The
e-Flow Audit: An Evaluation of Knowledge Flow
Within and Outside a High-tech Firm", Journal of
Knowledge Management, vol. 7, no. 1, pp. 6-19.
Tedmori, S., Jackson, T.W., Bouchlaghem, N.M. &
Nagaraju, R. 2006, "Expertise Profiling: Is Email
Used to Generate, Organise, Share, or Leverage
Knowledge", , eds. H. Rivard, E. Miresco & H.
Melham, , pp. 179.
Swaak, J., de Jong, T. & van Joolingen, W.R. 2004,
"The effects of discovery learning and expository
instruction on the acquisition of definitional and
intuitive knowledge", Journal of COmputer Assisted
Learning, vol. 20, no. 4, pp. 225-234.
Campbell, C.S., Maglio, P.P., Cozzi, A. & Dom, B.
2003, "Expertise identification using email
communications", twelfth international conference
on Information and knowledge managementNew
Orleans, LA, pp. 528.
Shanteau, J. & Stewart, T.R. 1992, "Why study
expert decision making? Some historical
perspectives and comments", Organizational
Behavior and Human Decision Processes, vol. 53,
no. 2, pp. 95-106.
Klimt, B. & Yang, Y. 2004, "Introducing the Enron
Corpus", First Conference on Email and Anti-spam.
Markus, M.L. 2001, "Toward a Theory of
Knowledge Reuse: Types of Knowledge Reuse
Situations and Factors in Reuse Success", Journal
of Management Information Systems, vol. 18, no. 1,
pp. 57-93.
Lichtenstein, S., Tedmori, S. and Jackson, T.W.,
2008, ''Socio-ethical issues for expertise location
from electronic mail'', International Journal of
Knowledge and Learning, 4(1), 58-74.
Lin, C., Cao, N., Liu, S., Papadimitriou, S., Sun, J.,
and Yan, X. 2009, “SmallBlue:Social Network
Analysis for Expertise Search and Collective
Intelligence”, IEEE 25th International Conference
on Data Engineering, pp. 1483-1486.
Thomas W. Jackson - is a Senior
Lecturer in the Department of
Information Science at Loughborough
University. Nicknamed ‘Dr. Email’ by
the media Tom and his research team
work in two main research areas,
Electronic
Communication
and
Information Retrieval within the
Workplace, and Applied and Theory
based Knowledge Management. He has published more than 70
papers in peer reviewed journals and conferences. He is on a
number of editorial boards for international journals and
reviews for many more. He has given a number of keynote talks
throughout the world. In both research fields Tom has, and
continues to work closely with both private and public sector
organisations throughout the world and over the last few years
he has brought in over £1M in research funding from research
councils, including EPSRC. He is currently working on many
research projects, including ones with The National Archives
and the Welsh Assembly Government surrounding information
management issues.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Sara Tedmori - In 2001, she received her
BSc degree in Computer Science from the
American University of Beirut, Lebanon.
In 2003, she obtained her MSc degree in
Multimedia and Internet Computing from
Loughborough University. In 2008, she
received her Engineering Doctorate in
Computer Science from Loughborough
University, UK. Currently she is
appointed as an assistant Professor in the Computer Science
Department at Princess Sumaya University of Technology,
Jordan. Her research interests include: Object Tracking, image
processing, expertise locator, knowledge extraction, knowledge
sharing, and privacy.
Chris
Hinde
is
Professor
of
Computational Intelligence in the
Department of Computer Science at
Loughborough University. His interests
are in various areas of Computational
Intelligence including fuzzy systems,
evolutionary computing, neural networks
and data mining. In particular he has
been working on contradictory and
inconsistent logics with a view to using them for data mining. A
recently completed project was concerned with railway
scheduling using an evolutionary system. He has been funded
by various research bodies, including EPSRC, for most of his
career and is a member of the EPSRC peer review college.
Amongst other activities he has examined over 100 PhD's.
Anoud Bani-Hani is an EngD research
Student at Loughborough University,
researching
into
Knowledge
Management in SME’s in the UK, with
specific focus on implementing an ERP
system into a low-tech SME. Prior to
joining the EngD scheme Anoud was a
Lecturer at Jordan University of Science
and Technology and holds an
undergraduate degree in Computer science and information
technology system from the same university and a Master
degree in Multimedia and Internet computing from
Loughborough University.
© 2012 ACADEMY PUBLISHER
127
128
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
MALURLS: A Lightweight Malicious Website
Classification Based on URL Features
Monther Aldwairi
Department of Network Engineering and Security, Jordan University of Science and Technology, Irbid, Jordan
munzer@just.edu.jo
Rami Alsalman
Department of Computer Engineering, Jordan University of Science and Technology, Irbid, Jordan
rsalsalman08@cit.just.edu.jo
Abstract—Surfing the World Wide Web (WWW) is
becoming a dangerous everyday task with the Web
becoming rich in all sorts of attacks. Websites are a major
source of many scams, phishing attacks, identity theft,
SPAM commerce and malwares. However, browsers,
blacklists and popup blockers are not enough to protect
users. That requires fast and accurate systems with the
ability to detect new malicious content. We propose a
lightweight system to detect malicious websites online based
on URL lexical and host features and call it MALURLs. The
system relies on Naïve Bayes classifier as a probabilistic
model to detect if the target website is a malicious or benign.
It introduces new features and employs self learning using
Genetic Algorithm to improve the classification speed and
precision. A small dataset is collected and expanded through
GA mutations to learn the system over short time and with
low memory usage. A completely independent testing
dataset is automatically gathered and verified using
different trusted web sources. They algorithm achieves an
average precision of 87%.
Index Terms— malicious websites, machine learning, genetic
algorithm, classification
I. INTRODUCTION
Internet access is an integral part of the modern life
and employees in today’s fast economy depend on
Internet connected smart phones, laptops and personal
assistants to perform their jobs on the go. Even regular
Joe and young children are becoming techsavvy and
cannot live without Internet. Users shop, check movies,
bank accounts, email, health insurance, they renew
driving licenses, pay bills, make calls over IP, chat and
play games, just to name a few daily activities. Such
Internet access takes place through web browsers making
them the most popular application for most users.
As browser vendors race to introduce more features
and new functionalities more vulnerabilities arise and
more personal data are put at risk. Browsers have become
the main system entry point for many attacks that aims at
stealing private data and manipulating users to reveal
sensitive information. Unsuspecting web surfers are not
aware of the many drive-by-downloads of malwares, ad
wares, spywares and Trojans to their devices. Just a
single visit to a shady website is sufficient to allow the
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.128-133
intruder to detect vulnerabilities in the surfer’s computer
and inject a malware that might enable the intruder to
gain remote access or open a backdoor for future
blunders.
Users do not have to visit pornographic or hacker
websites to get compromised. Commerce related SPAM
such as pharmaceuticals and fake products are one way to
coerce users to click and access malicious websites. In
addition, they can be redirected to such websites through
more organized schemes such as fast flux networks
(FFN) [1]. Consequently, users are easily tricked to
reveal private information using phishing and pharming
attacks [2]. In addition to all of that, browsers collect
sensitive data such as favorites, cache files, history file,
cookies, form data and passwords. This puts such
information at risk and keeping your browser and
computer up to date will not cut it. For instance, cache
timer sniffing enables intruders to determine websites you
have visited.
Finding and identifying such websites is no simple task
due to the ever growing World Wide Web and the
dynamic nature of malicious websites. Blacklisting
services rose to the challenge and were encapsulated into
browsers, toolbars and search engines. The lists are
constructed through manual reporting, honeypots or web
spiders. But blacklists grow uncontrollably and become a
performance bottleneck. Incorrect listing is a major
problem, due to reporting, analysis and record keeping
mistakes. Therefore, legitimate websites may be
incorrectly evaluated and listed, while malicious websites
are not listed because they are new and haven’t been
analyzed yet. Researchers have been very active in
devising online and offline solutions to classify malicious
websites and make web surfing safer. Shortly we give a
brief survey of the current state of art techniques for
classifying websites.
In this paper we propose lightweight statistical selflearning scheme to classify websites based on their
features. It is fast and designed to run online to protect
the users. We use a Naïve Bayes classifier to classify the
websites into two classes: malicious or benign. The
number of features used is small and they fall under one
of three categories: lexical, host-based or special features.
Features include those suggested by McGrath et al. [3]
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
and Ma et al. [4]. We add special features to improve the
classification
accuracy
such
as
JavaScript
Enable/Disable, Document Frequency, and Title Tag. In
addition, Genetic Algorithm is used to expand the
training dataset through mutations to learn the Naïve
Bayes classifier better and faster without the need to deal
with huge datasets. The authors presented preliminary
results in a previous paper [5] and in this paper they
expand the original work by including more
implementation details and adding additional results.
The rest of the paper is organized as follows. Section II
surveys the related work. Section III details the
methodology followed to classify websites, this includes
the list features, collecting the training and testing
datasets and how the Naïve Bayes and GA are used.
Section IV presents the experimental results.
II. RELATED WORK
Blacklisting was and still is a popular technique.
Whittaker et al. [6] offline analyzed millions of pages
daily from the noisy Google's phishing blacklist. Their
main contribution was achieving 90% classification
accuracy for phishing pages after a three weeks training.
PhishNet [7] used approximate pattern matching
algorithm to match URL components against blacklist
entries. Though the above techniques tried to
automatically manage blacklists and increase their
accuracy, they are still insufficient and suffer from their
growing size and incorrect listing. Blacklists can be
combined with other techniques that uses machine
learning to classify malicious websites.
One of the earliest classification systems for malicious
websites was concerned with the detection of SPAM in
blog posts. Blog identification and splog detection by
Kolari et al. [8] used the activity and comments generated
by a blog post as the main classification feature in
addition to ping update services. Support Vector
Machines (SVM) was used with only linear kernel in all
experiments and reported moderate results. Subsequent
work focused on detecting Phishing URLs in SPAM
emails. Garera et al. [9] main contribution was
identifying eighteen features to detect phishing URL
embedded in SPAM. They used linear regression
compare millions of Google’s toolbar URLs to identify
777 phishing pages a day and 9% of the users that visit
them are potential victims. McGrath et al. [3] studied
phishing infrastructure and the anatomy of phishing
URLs. They pointed out the importance of features such
as the URL length, linked-to domains age, number of
links in e-mails and the number of dots in the URL.
PhishDef [10] used features that resist obfuscation and
suggested used the AROW algorithm to achieve higher
accuracy.
To further increase the accuracy several approaches
focused on page content statistics such as Seifert et al.
[11]. They added features derived from JavaScript and
HTML tags such as redirects, long script code lines and
shell code. Seifert et al. used features from the page
contents such as the number of HTML script tags and
size of iframe tags. Cova [12] et al. went too far by
© 2012 ACADEMY PUBLISHER
129
profiling the normal JavaScript behavior and applying
anomaly detection which is prone to high false positives.
Anomaly detecting works by extracting features during
the normal learning phase based on a specific model. In
the testing phase the new feature values for the websites
to be tested are checked against the training models
representing the normal behavior. The features used
include: the number of code executions, code length,
number of bytes, shell codes and the difference in
returned pages for different browsers and the number of
redirections. The Prophiler by Canali [13] used HTML
tag counts, percentage of the JavaScript code in the page,
percentage of whitespace, entropy of the script, entropy
of the of the strings declared, number of embed tags,
presence of meta refresh tags, the number of elements
whose source is on an external domain and the number of
characters in the page. While improving accuracy the
Prophiler significantly increased the number of features
to eighty eight. In addition to the increased overhead due
to the statistically processing the page content, those
techniques suffered from the inherent danger of having to
access the malicious page and download the content
before deciding it was malicious.
Ma et al. [4] used Yahoo−PhishTank dataset and
validated their work using three machine learning models
Naïve Bayes, SVM with an RBF kernel and regularized
logistic regression. Later, Ma et al. [14] [15] developed a
light weight algorithm for website classification based on
lexical and host-based features while excluding page
properties. It was designed as real-time, low-cost and fast
alternatives for black listing. They reported 3.5% error
rates and 10–15% false negatives, but the tradeoff was
between memory usage and accuracy. However, the main
disadvantage was the fact that they use tens and hundreds
of thousands of features to achieve their results. Another
disadvantage shared among all previous approaches is
collecting and handling a large number of websites and
features which makes it hard to run them online.
III. MTHODOLOGY
To address the drawback of previous wok we need to
identify malicious websites. We define a malicious
webpage as a page that downloads a file, uploads a file,
collects data, installs an application, opens a pop
window(s), displays an advertisement or any combination
of the above without the knowledge or consent of the
user. We manually construct our own dataset and label
websites as either benign or malicious based on trusted
web directories. The complete MALURLs framework is
shown in Fig. 1. In step 3 the features for the training
dataset are calculated through various sources. We collect
100 benign and 100 malicious sites. In steps 5 and 6 the
dataset is expanded to 10000 records using Genetic
Algorithm (GA). GA self learns the classifier through
mutations on the dataset. Based on a fitness function we
can use mutations and crossovers from the current dataset
to generate a larger dataset and grantee not to learn our
classifier based on specific domain. In step 7 Naïve
Bayes are trained using part of the collected features.
Finally, a completely different dataset of 200 websites is
130
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Figure 1. MALURLs framework.
used for testing. The testing dataset is collected and
classified automatically into benign and malicious. The
features are calculated based on a same web sources used
in training. It is worth mentioning that the testing and
training datasets are completely independent and are
verified and classified using different trusted web sources
to eliminate any chance of data poisoning. Unlike most of
previous approaches which used the same source, mainly
PhishTank.
The following Subsections explain the three feature
groups and the basis for their selection, how the training
dataset is collected, how the Naïve Bayes classifier
works, the Genetic Algorithm and finally how the testing
dataset is built.
A. Features
Features used fall into one of three categories: lexical,
host-based features and special features.
1) Lexical Features
URL stands for uniform resource locator or formerly
the universal resource locator. URL and uniform resource
identifier (URI) are equivalent and identify any document
retrieved over the WWW. The URL has three main parts:
the protocol, hostname and path. Consider the following
URL for example: “http://www.just.edu.jo/~munzer/
Courses/INCS741/Lec/ch3.ppt”. The protocol is: http://”,
the hostname is: “www.just.edu.jo” and the path is:
“~munzer/Courses/ INCS741/Lec/ch3.ppt”.
Lexical features are the properties of the URL itself
and do not include content of the page it points to. The
URL properties include the length of the top level domain
(TLD), other domains, the hostname, URL length, as well
as the number of dots in the URL. In addition, lexical
features include each token in the hostname (delimited by
‘.’) and tokens in the path URL delimited by ‘/’, ‘?’, ‘+’,
‘ ̶ ’, ‘%’, ‘&’, ‘.’, ‘=’, and ‘_’. Those feature groups are
known as a “bag-of-words”. The features above tell a lot
© 2012 ACADEMY PUBLISHER
about a webpage. The domain might indicate a
blacklisted malicious content provider. A large number of
NULL token probably attributed to too many slashes
might indicate an http denial of service attack (DOS) on
Microsoft Internet Information Server (IIS).
2) Host-based Features
Host-based features are derived from the host
properties such as the IP address, geographic properties,
domain name properties, DNS time to live (TTL), DNS
A, DNS PTR and DNS MX records as well as WHOIS
information and dates. Those features are very important
and can help any classifier better the detection process.
They help address a lot of important questions such as:
does the IP address belong to a geographical location
associated with malicious content? Does the PTR record
resolve an IP that belongs to the host? Do the IP
addresses of the DNS records belong to one autonomous
system (AS)?
3) Special Features
Those are new features that do not fit under the
aforementioned categories or reported good results with
previous systems. Some are simple to get a value for such
as JS Enable/Disable, HTML Title tag content
(<title></title>), 3-4-5 grams (n-grams) and Term
Frequency and Inverse Document Frequency (TF-IDF).
JavaScript code usually is downloaded and run on the
client’s browser which can be very dangerous. Term
frequency is the number of times a term occurs in a
document. The inverse document frequency is the
logarithm of the number of documents divided by the
number of documents containing the term and it measures
the importance of a term. TF-IDF is commonly used in
search engines, classification and data mining and finally
3-4-5 grams take longer to calculate than the other
features.
Other features require significant computation time
such as Anchors or bag-of-anchors which are extracted
from all URLs in Anchor tags on the page being
examined. Table I below shows a list of features’ groups
used for training and testing purposes.
B. Training Dataset
The dataset is composed to 200 websites, half benign
and the other half is malicious. The websites are chosen
randomly and ALEXA Web Information Company’s
TABLE I.
FEATURE GROUPS
JS-Enable-Disable
Document
Frequency DF
Title tag
<title>??</title>
3-4-5 grams
TF-IDF weighting
Blacklists
WHOIS dates
IP address misc
Lexical misc
4grams
Features
DNS PTR record
WHOIS info
Connection
speed
TLD + domain
DNS A record
Geographic
Hostname
Words+URLs
Meta+link
URLs+anchors
URLs+anchors+meta
Path tokens
Last token of the
path
Spamassassin
plugin
TLD
DNS TTL
DNS MX record
Bag-of-words
URLs
Anchors
Meta tags
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
website [16] is used to determine benign websites while
PhishTank dataset [17] is used to determine malicious
websites. ALEXA is one of the most influential and
trusted WWW information companies. It provides
information about websites including Internet traffic and
top sites rankings. MALURLs uses IP tracer website [18]
to extract the URL, DNS, IP address and the geographic
properties for benign websites.
PhishTank is an open source anti-phishing website that
is widely used by almost all major browsers and WWW
vendors such as Mozilla, Yahoo and McAfee. It offers
phish pages verification tool through a voting system and
issues annual phishing reports. PhishTank dataset
contains partial information about malicious websites
such as URL, DNS, IP-address and the geographic
information. However, not all features are available on
PhishTank, particularly the new special features.
Therefore, we use Sphider [19] which is an open source
web spider and a search engine. Sphider performs full
text indexing, for example it finds links anywhere in a
document whether in href or even in JavaScript strings.
Sphider is able to calculate the Term Frequency (TF),
Document Frequency (DF) and Inverse Document
Frequency (IDF) features.
C. Naïve Bayes classifier
Bayes [20] is a probabilistic model based on Bayesian
theorem. Though it is simple but often outperforms most
of the other classifiers especially if trained using
supervised learning methods. Naïve Bayes classifiers
assume that the effect on a class from a feature is
independent of the values of other features. This
conditional independence simplifies the computation
without sacrificing the accuracy. This makes Naïve Bayes
a perfect match for our lightweight online algorithm.
Bayesian theorem for example, calculates the probability
that a website is malicious from the independent
probabilities that a website is from a geographic location
that generates fake traffic, has a random and very long
hostname, has an IP address that does not match the DNS
A recor and so on.
In our case the number of features and their values
range are large. If C represents the class and F represents
a feature then the conditional probability (Pr) of C given
F is calculated according to (1).
( | )=
( | )
( )
( )
(1)
D. Genetic Algorithm
To generate a larger dataset from the initial dataset we
use Genetic Algorithm. GA is a biologically inspired
algorithm that applies the principles of evolution and
natural selection. The algorithms starts with an initial
population encoded as a chromosome structure which is
composed of genes encoded as numbers or characters. In
our case the initial population represents the group of
features for the training dataset. Chromosome goodness is
evaluated using a fitness function that uses mutations and
crossovers to simulate the mutation of species. The fittest
chromosomes are selected and the process is repeated till
we converge to a solution [21].
© 2012 ACADEMY PUBLISHER
131
The initial population is the dataset collected as
specified in Subsection B and used in the learning phase.
Mutations are applied on the initial dataset which is
composed of a number of features or genes. Mutations
are simply applying changes to certain features such as
changing JS-Enable-Disabled from True to False (binary
encoding) or to add a random amount between 0.2-0.3 to
DF and TF (value encoding).
In the testing step the fitness function is calculated by
multiplying the probability values for the all features as
shown in (2). The fitness function is used to calculate
malicious and benign probabilities. The website is
classified based on the highest probability.
(
,
ℎ ,
,
,
( )=∏
,
−
,
ℎ
,
,
,
)
,
(2)
E. Testing Dataset
For testing we collect 200 URLs (100 malicious, 100
benign) using WOT Mozilla Plug-in [22]. The features
values are collected the same way as in the training
dataset. WOT is a traffic-light style rating system where
green means (benign) and red means stop (malicious).
The rating of a website depends on a combination of user
ratings and data from trusted sources such as Malware
Patrol [23], Panda [24], PhishTank, TRUSTe [25],
hpHosts [26] and SpamCop [27]. In addition, WOT
enables users to evaluate the trustworthiness of a website
and incorporate their ratings in calculating the reputation
of a website.
IV. IMPLEMENTATION AND RESULTS
We implement MALURLs using PHP programming
language and MySQL database. Equation (3) defines the
precision metric used to evaluate the relative accuracy of
MALURLs using different features. Precision (P) is a
measure of the usefulness of the retrieved documents.
=
(3)
The testing dataset of 200 instances was divided into
five different subsets and the average precision with and
without Genetic Algorithm was calculated as shown in
Table II. The use of GA to expand the training dataset
results in a significant improvement in the classification
precision. The average precision for classifying a website
as benign or malicious using GA is 87% when using all
feature groups.
We run experiments to measure the improvement in
TABLE II.
OVERALL PRECISION WITH AND WITHOUT GA
Dataset
Number
1
2
3
4
5
Average
Precision (%)
Without GA
With GA
70
75
80
80
65
74
85
90
95
80
85
87
132
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
TABLE III.
OVERALL PRECISION WITH AND WITHOUT TF-IDF
Dataset
Number
1
2
3
4
5
Average
Precision using GA (%)
Without JS
With JS
60
65
90
55
60
66
70
60
95
80
75
76
TABLE IV.
PRECISION WITH AND WITHOUT JS-ENABLE-DISABLE
Dataset
Number
1
2
3
4
5
Average
Precision using GA (%)
Without JS
With JS
55
50
75
70
65
63
65
60
75
65
80
69
TABLE V.
PRECISION WITH AND WITHOUT 3-4-5 GRAMS
Dataset
Number
1
2
3
4
5
Average
Precision using GA (%)
Without n-grams
With n-grams
80
70
85
60
90
77
80
70
85
75
90
80
precision attained by adding the new individual features
to MALURLs. The features added include TF-IDF, JSEnable-Disable and 3-4-5 grams. The addition of TF-IDF
results in a significant increase in classification precision
from 66% to 76% as shown in Table III. This is expected
because of the volume of information presented by this
feature. Adding the JS-Enable-Disable did show a very
good increase in average precision from 63% to 69% as
illustrated by Table IV. The experiments to measure the
improvement in MALURLs precision achieved by adding
3-4-5 grams show a small increase in average precision
from 77% to 80% as illustrated by Table V. 3-4-5 grams
calculation is complex, takes a long time and puts the
user at risk due to the need to download the document.
Therefore n-grams can be deemed irrelevant because of
the high overhead and small improvement which is
consistent with our goal of keeping the algorithm light
weight.
V. CONCLUSIONS
This paper presents a new website classification
system based on URL, host-based and special feature. We
experiment with various features and determine the ones
that improve the precision with minimum overhead. The
data is collected using WOT Mozilla plug-in and the
features are calculated using various web resources.
MALURLs system reduces the training time using GA to
expand the training dataset and learn the Naïve Bayes
classifier. The experimental results show the average
system precision of 87%. The additional features proved
© 2012 ACADEMY PUBLISHER
valuable to improve the overall classification precision.
TF-IDF improved precision by up to 10%, JS-EnableDisable improvement was about 6% while 3-4-5 grams
improvement was limited to 3%.
REFERENCES
[1] A. Caglayan, M. Toothaker, D. Drapeau, D. Burke and G.
Eaton, “Real-time Detection and Classification of Fast
Flux Service Networks”, In Proceedings of the
Cybersecurity Applications and Technology Conference
for Homeland Security (CATCH), Washington, DC, Mar
2009.
[2] J. Zdziarski, W. Yang, P .Judge, “Approaches to Phishing
Identification using Match and Probabilistic Digital
Fingerprinting Techniques”, In Proceedings of the MIT
Spam Conference, 2006.
[3] D. McGrath and M. Gupta, “Behind Phishing: An
Examination of Phisher Modi Operandi”, In Proceedings of
the USENIX Workshop on Large-scale Exploits and
Emergent Threats (LEET), San Fransicso, CA, Apr 2008.
[4] J. Ma, L. Saul, S. Savage, and G. Voelker, “Beyond
Blacklists: Learning to Detect Malicious Websites from
Suspicious URLs”, In Proceedings of the ACM SIGKDD
Conference, Paris, France, Jun 2009.
[5] M. Aldwairi, R. Alsalman. “MALURLs: Malicious URLs
Classification System”. In Proceedings of the Annual
International Conference on Information Theory and
Applications (ITA), Singapore, Feb 2011.
[6] C. Whittaker, B. Ryner, and M. Nazif, “Large-Scale
Automatic Classification of Phishing Pages”, In
Proceedings of the 17th Annual Network and Distributed
System Security Symposium (NDSS’10), San Diego, CA,
Mar 2010.
[7] P. Prakash, M. Kumar, R. R. Kompella and M. Gupta,
“PhishNet: Predictive Blacklisting to Detect Phishing
Attacks”, In Proceedings of INFOCOM ’10, San Diego,
California, Mar 2010.
[8] P. Kolari, T. Finin, and A. Joshi, “SVMs for the
Blogosphere: Blog Identification and Splog Detection”, In
AAAI Spring Symposium on Computational Approaches
to Analyzing Weblogs, Mar 2006.
[9] S. Garera, N. Provos, M. Chew, and A. D. Rubin, “A
Framework for Detection and Measurement of Phishing
Attacks”, In Proceedings of the ACM Workshop on Rapid
Malcode (WORM), Alexandria, VA, Nov 2007.
[10] A. Le, A. Markopoulou and M. Faloutsos, “PhishDef: URL
Names Say It All”, In Proceedings of the 30th IEEE
INFOCOM 2011 (Mini Conference), Shanghai, China,
April 10-15, 2011.
[11] C. Seifert, I. Welch and P. Komisarczuk, “Identification of
Malicious Web Pages with Static Heuristics”, In
Proceedings of the Australasian Telecommunication
Networks and Applications Conference (ATNAC), 2008.
[12] M. Cova, C. Kruegel and G. Vigna, "Detection and
Analysis of drive-by-Download Attacks and Malicious
JavaScript Code", In Proceedings of the 19th International
Conference on World Wide Web (WWW'10), Raleigh,
NC, Apr 2010.
[13] D. Canali, M. Cova, G. Vigna and C. Kruegel, “Prophiler:
a Fast Filter for the Large-Scale Detection of Malicious
Web Pages", In Proceedings of the 20th International
World Wide Web Conference (WWW), Hyderabad, India,
Mar 2011.
[14] J. Ma, L. Saul, S. Savage, and G. Voelker, “Identifying
Suspicious URLs: An Application of Large-Scale Online
Learning”, In Proceedings of the International Conference
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
on Machine Learning (ICML), Montreal, Quebec, Jun
2009.
J. Ma, L. Saul, S. Savage, and G. Voelker, “Learning to
Detect Malicious URLs”, ACM Transactions on Intelligent
Systems, vol. 2, no. 3, Apr 2011.
ALEXA Web Information Company’s website,
http://www.alexa.com/, last access Feb 2011.
PhishTank, http://www.phishtank.com/, last access Feb
2011.
IP Tracer Website. http://www.ip-adress.com/ip_tracer/,
last access Feb 2011.
Sphider, http://www.sphider.eu/, last access Feb 2011
G. John, G. and P. Langley, Estimating continuous
distributions in Bayesian classifiers, In proceedings of the
Eleventh Conference on Uncertainty in Artificial
Intelligence (pp. 338–345), San Francisco, CA, 1995.
J. Koza, M. Keane, M. Streeter, W. Mydlowec, J. Yu and
G. Lanza “Genetic Programming IV: Routine HumanCompetitive Machine Intelligence”, Kluwer Academic
Publishers, 2003.
Web
of
Trust
Mozilla
Plug-in,
http://www.mywot.com/en/download, last access Feb
2011.
Malware Patrol. http://www.malware.com.br/, last access
Feb 2011.
Panda WOT. http://www.pandasecurity.com/, last access
Feb 2011.
TRUSTe – WOT. http://www.truste.org/, last access Feb
2011.
hpHosts online. http://www.hosts-file.net/, last access Feb
2011.
SpamCop. http://www.spamcop.net/, last access Feb 2011.
Monther Aldwairi was born in Irbid,
Jordan in 1976. Aldwairi received a
B.S. in electrical engineering from
Jordan University of Science and
University (JUST) in 1998, and his
M.S. and PhD in computer engineering
from North Carolina State University
© 2012 ACADEMY PUBLISHER
133
(NCSU), Raleigh, NC in 2001 and 2006, respectively.
He is an Assistant Professor of Network Engineering and
Security Department at Jordan University of Science and
Technology, where he has been since 2007. He is also the Vice
Dean of Faculty of Computer and Information Technology since
2010 and was the Assistant Dean for Student Affairs in 2009. In
addition, he is an Adjunct Professor at New York Institute of
Technology (NYiT) since 2009. He worked as Post-Doctoral
Research Associate in 2007 and as a research assistant at NCSU
from 2001 to 2006. He interned at Borland Software
Corporation in 2001. He worked as a system integration
engineer for ARAMEX from 1998 to 2000. His research
interests are in network and web security, intrusion detection
and forensics, artificial intelligence, pattern matching, natural
language processing and bioinformatics. He published several
well cited articles.
Dr. Aldwairi is an IEEE and Jordan association of engineers’
member. He served at the steering and TPC committees of
renowned conferences and he is a reviewer for several
periodicals. He organized the Imagine cup 2011-Jordan and the
national technology parade 2010.
Rami Alsalman was born in Irbid, Jordan
in 1986. He received his B.S. in computer
information
systems
from
Jordan
University of Science and University in
2008. He received his M.S. degree in
computer engineering JUST in 2011.
He is currently working on his PhD in
cognitive systems from the computer
science department at the University of
Bremen, Germany. He has three
publications in AI and security areas.
He received a three months research scholarship at the
Heinrich-Heine University of Düsseldorf, Germany during the
summer of 2010. His research interests include security, data
mining, information assurance and data reasoning.
Eng Alsalman is an editorial board at an International
Journal of Web Applications. Additionally, he was a committee
member in International Conference on Informatics,
Cybernetics, and Computer Applications (ICICCA2010).
134
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Vision-based Presentation Modeling of Web
Applications: A Reverse Engineering Approach
Natheer Khasawneh
Department of Software Engineering, Jordan University of Science and Technology, Irbid, Jordan
Email: natheer@just.edu.jo
Oduy Samarah
Department of Computer Engineering, Jordan University of Science and Technology, Irbid, Jordan
Email: oasamarah09@cit.just.edu.jo
Safwan Al-Omari
Department of Software Engineering, Jordan University of Science and Technology, Irbid, Jordan
Email: ssomari@just.edu.jo
Stefan Conrad
Institute of Computer Science, Heinrich Heine University, Dusseldorf, Germany
Email: conrad@cs.uni-duesseldorf.de
Abstract—Presentation modeling, which captures the layout
of an HTML page, is a very important aspect of modeling
Web Applications (WAs). However, presentation modeling
is often neglected during forward engineering of Web
Applications; therefore, most of these applications are
poorly modeled or not modeled at all. This paper discusses
the design, implementation, and evaluation of a reverse
engineering tool that extracts and builds appropriate UML
presentation model of existing Web Applications. The tool
consists of three steps. First, we identify and extract visual
blocks and presentation elements of an HTML page such as
navigation bars, header sections, text input, etc. In this step,
we adopt the VIPS algorithm, which divides an HTML into
semantically coherent blocks. Second, the identified
presentation elements in step one are mapped to the most
appropriate UML presentation model elements. Third, the
resulting presentation model is made available in
Magicdraw for manipulation. Our approach is applied and
evaluated in the Goalzz home page.
Index Terms—Reverse Engineering, Web Application, Web
UML, Vision-based Page Segmentation
I. INTRODUCTION
Recently, many applications and services have evolved
from being stand-alone and monolithic applications into
web applications. According to a recent study [1], most of
the existing web applications lack proper modeling,
which is necessary for maintenance, reengineering, and
proper evolution to emerging web technologies. For the
purpose of modeling legacy web applications, there is an
urgent need to have a reverse engineering method to
extract models of existing web applications. Chikofsky
describes reverse engineering as “the process of
analyzing a subject system to identify the system’s
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.134-141
components and their interrelationships and create
representations of the system in another form or at a
higher level of abstraction” [2].
Current modeling languages and methodologies are not
sufficient for capturing all aspects of web applications.
UML for example is not sufficient to express the
hyperlinks between different HTML pages. Modeling of
web applications can be performed at three levels:
1. Content modeling: focuses on modeling data in an
HTML page.
2. Hyper text modeling: focuses on modeling links
between HTML pages in a web application.
3. Presentation modeling: focuses on the layout of the
items inside a particular HTML page.
In this paper we focus on the presentation model,
which is used to model the page layout in a UML
presentation model. The approach we follow in
generating a presentation model is based on page
segmentation method discussed in [3]. Page segmentation
divides the page into different blocks according to its
visual appearance when rendered in a web browser.
The reset of the paper is organized as follows: in
Section 2, we provide an overview of the related work;
Section 3 introduces the proposed approach in details;
Implementation details are discussed in Section 4;
Section 5 discusses case studies to illustrate and
demonstrate our approach; finally in Section 6, we sketch
concluding remarks and future work.
II. RELATED WORK
In the literature there are many methods and tools of
web reverse engineering which are built on the standard
of reverse engineering techniques. These methods and
tools can be used to describe and model the web
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
applications with respect to different levels: content,
hypertext, and presentation [4].
The UML modeling language is the most widely used
during the forward engineering design process and in
many web application reverse engineering techniques.
For example, in [1] authors defined a process for reverse
engineering by describing a method for understanding
web applications to be easily maintained and evolved. In
[5], authors showed how web applications with UML
presentation can be easily maintained. The approach in
[6] is based on structured and model based techniques. In
this approach the HTML page is divided into several
blocks according to a cognitive visual analysis. After that
the specific patterns with these blocks are extracted to
produce structural blocks and through these structural
blocks a conceptual model is represented. The approach
in [7] relies on HTML pages analysis by extracting the
useful information from the web page and analyze the
extracted information using the domain ontology and
form the analysis results the UML conceptual schema is
generated.
III. APPROACH AND METHODOLOGY
In this paper, we present a new approach to reverse
engineer existing web applications into UML
presentation model. The proposed approach focuses on
discovering the structure of the web page and presenting
the structure in UML presentation model, as shown in
Fig. 1.
135
model generation stage. Page segmentation stage accepts
an HTML page as input and produces an XML
description of segmented page. Mapping process stage
transforms the XML description to Web UML XML
description. Finally, UML generation Stage accepts a
Web UML XML description of the segmented page and
outputs the UML presentation model.
In the rest of this section, we describe, in details, page
segmentation stage, web UML XML, mapping process
stage, and UML model generation stage, respectively.
A. Page Segmentation Stage
In this stage we use Vision-based Page Segmentation
Algorithm (VIPS) [3] to segment the web page into
different blocks. VIPS incorporates both the page
appearance (visual cues) and Document Object Model
(DOM) to perform the segmentation.
VIPS is done in three steps (Fig. 2): block extraction,
separator detection, and content structure construction.
These steps are repeated recursively several times until a
user-defined threshold is reached. The threshold is called
Permitted Degree of Coherence (PDoC) and it is based on
Degree of Coherence (DoC). DoC is a numeric value
between 1 and 10 that increases as the consistency
between blocks increases. Following is a description of
each step in VIPS:
Visual Block Extraction
The input to the visual block extraction is the visual
Webpages
Page Segmentation
Stage 1
VIPS XML Description of
Segmented Page
Mapping Process
Stage 2
Web UML XML Description of
Segmented Page
Parsing Web UML XML Description
by Magicdraw Tool
Stage 3
UML Presentation Model of
Segmented Webpage
Figure 1. Approach Architecture.
The presented method consists of three stages: page
segmentation stage, mapping process stage, and UML
© 2012 ACADEMY PUBLISHER
Figure 2. Flowchart of the Segmentation Process.
cues and the DOM tree of the web page. During this step
each node in the DOM tree is matched with the block
where it belongs to in the visual cues. From the tree node,
136
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
VIPS starts looking if the sub nodes belong to the same
block. Nodes with a coherence value less than PDoC are
matched together to belong to the same group. This
process is repeated recursively for the sub roots of the
unmatched nodes until all nodes are matched to a block in
the visual cues.
Visual Separator Detection
The input to this step is the collection of the extracted
blocks from the previous step, whereas, the output is a set
of separators which separate different blocks. There are
two types of separators, horizontal separator and vertical
separator. The process starts by one separator which
spans the whole page. Then blocks are added one by one
and the separator gets updated according to the following
rules:
1. If the added block falls inside the separator, the
separator will be splitted into two. The splitting will
be done either vertically or horizontally.
2. If the added block covers part of the separator the
separator will be resized.
3. If the added block covers the whole separator, the
separator will be removed.
Weight separator is assigned to each separator by
considering factors that show how similar the
neighboring blocks are. For a separator which separates
two blocks, the more the two blocks differ the higher the
weight that will be assigned for the separator and vice
versa. The used factors are: distance between blocks,
overlapping with HTML tags, background differences,
font differences, and structure similarity.
Content Structure Construction
In this step content structure is constructed by merging
lowest weight separators with the neighboring separators.
The newly merged separator will be given a DoC value
equals to the maximum DoC of the merged separators.
This step is iterative until a separator with maximum
weight is reached. Finally, each node in the newly
generated block is checked whether its DoC meets the
condition given by PDoC or not. If not, the Visual
Extraction Process starts over. Otherwise, the process
terminates and the page structure is generated in VIPS
XML description format.
VIPS XML description format is the immediate output
of the VIPS algorithm. It is very simple and
understandable description, which is defined for blocks
and their content in the segmented page. The VIPS XML
description captures and describes some attributes for
each block and HTML elements in a web page, as shown
in Table I. This table contains the main attributes which
are used during the mapping process stage.
Fig. 3 represents the VIPS XML description for blocks
with ID (1-2) which contains three elements irrespective
of the type of these elements (block or primitive HTML
Element). Fig. 3 also shows more information for block
such as coordinates and not containing any tables or
images. In contrast, Fig. 4 represents VIPS XML
description for primitive HTML Element (anchor) with
ID (1-2-1), and also Fig. 4 shows more information about
© 2012 ACADEMY PUBLISHER
TABLE I.
VIPS XML ATTRIBUTES AND THEIR DESCRIPTION
VIPS XML
Attributes
ContainImg
IsImg
Descriptions
Determine whether the block contains
image
Determine whether the HTML Element
is image
ContainTable
Determine whether the block contains
table
ContainP
Determine whether the HTML Element
contains text
TextLen
Determine the length of text in blokes
or HTML Elements
DOMCldNum
Determine whether the element is block
or primitive HTML tag ( it is block if
DOMCldNum >1)
ObjectRectLeft
ObjectRectTop
ObjectRectWidth
ObjectRectHeight
Determine the coordinates for blocks or
HTML Elements
Content
Determine the content of blocks and
anchor which appears for user .(used for
data mining techniques )
SRC
Determine the source of image, anchor
or any HTML Element
ID
Unique ID for block or HTML Element,
which is used for specific purpose
during the implementation.
order
Unique ID (integer number)
the element such as coordinates and Uniform Resource
Locator (URL) for anchor in SRC attribute. Through ID
attribute, we can know that the block with ID (1-2)
contains the HTML element with ID (1-2-1).
<LayoutNode FrameSourceIndex="0" SourceIndex="16"
DoC="10" ContainImg="0" IsImg="false"
ContainTable="false" ContainP="0" TextLen="15"
LinkTextLen="15" DOMCldNum="3" FontSize="7.5"
FontWeight="700" BgColor="#00ffff"
ObjectRectLeft="10" ObjectRectTop="289"
ObjectRectWidth="202" ObjectRectHeight="236" ID="1-2"
order="10">
Figure 3. VIPS XML Description for Block.
<LayoutNode FrameSourceIndex="0" SourceIndex="18"
DoC="11" ContainImg="0" ContainTable="false"
ContainP="0" TextLen="5" LinkTextLen="5"
DOMCldNum="1" FontSize="7.5" FontWeight="700"
BgColor="#00ffff" IsImg="false" ObjectRectLeft="60"
ObjectRectTop="308" ObjectRectWidth="95"
ObjectRectHeight="36" Content="Page1 " SRC="&lt;A
href= &quot; FirstPage. html&quot; &gt;Page1&lt;/A&gt; "
ID="1-2-1" order="11"/>
Figure 4. VIPS XML Description for Anchor HTML Elements.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
B. Web UML XML
Web UML XML description is based on UML-Based
Web Engineering (UWE) Metamodel [8]. UWE is a
methodology used for Web application modeling
purpose, especially, structure modeling and behavior
modeling. Furthermore, this methodology provides
guidelines for systematic modeling of Web applications.
UWE comprises four main parts to model web
application: Notations, Methods, Metamodel, and
Process.
Next we present a brief description about the UWE
notation before discussing the process of transforming
VIPS XML description format to Web UML XML
format.
UWE Notation
For UML presentational model, UWE defines
stereotypes for user interface elements, these elements
are divided into two parts: primitive user interface
(UIElement) elements and user interface container
(UIContainer) elements which contain a collection of
primitive user interface elements.
A case-tool Magicdraw supports the design and
model of web applications with different aspects by using
UWE notations and Metamodels as plug-in. This casetool is used to parse the Web UML XML description to
generate UML presentation model in generation model
stage.
Table II summarizes the UWE Stereotypes, which are
used to model web application into UML presentation
model. These Stereotypes are:
1.
Page: A Page element is the area of the user interface
which contains all of UIElement elements and
UIContainer elements. The page constructs the root
of presentation model.
2. Anchor: An anchor element permits the user to move
from page to another page, or from location to
another location on the same page.
3. Group Presentation: A group presentation element is
used to define a set of UIElement elements, such as a
collection of anchors.
4. Text: A text element is used to display a sequence of
characters.
5. Textinput: A text input element permits the user to
enter text.
6. Form: A form element contains a collection of
UIElement elements that are used to provide data for
a submitted process.
7. Button: A button element allows the user to initiate
some actions on the web page. Actions include
submitting the content for Textinput element, playing
video, displaying image, triggering anchor and so on.
8. Selection: A selection element displays a list of items
for the user to select one or more items.
9. Image: An image element is used to display the
image.
10. Media Object: MediaObject elements are used to
play multimedia objects such audio and video.
137
TABLE II.
UWE SEROTYPES SYMBOL.
Stereotypes
Symobl
Page
Anchor
Text
Text input
Selection
File upload
Image
mediaObject
Form
Button
Collection
Custom component
11. File upload: A File upload element allows user to
upload files.
12. Custom component: UWE also defines custom
component stereotypes for custom HTML elements
which are not defined by UWE.
The Web UML XML description is a complicated
description which is used to describe the UWE
stereotypes in XML format. For that, this type of XML
description can be interpreted and rendered to UWE
stereotypes by any case-tool that is capable of parsing this
XML description to build a UML presentation model.
The Web UML XML description is divided into three
main parts, each part specifies some properties for
elements. Fig. 5, Fig. 6, and Fig. 7 contain the UML
XML description for a simple block of web page, this
block is represented by the DIV HTML tag which
consists of two Textinput elements and one submit button
as shown in Fig. 8.
The Web UML XML description in Fig. 5 specifies the
ID number and name for each element, and also
determines whether the element is visible or not in UML
presentation model. The benefit of visibility property
appears through modeling the hidden HTML elements
that are not displayed by the browser.
<packagedElement xmi:type='uml:Class' xmi:id='1_1_2'
name='Administrator Login' visibility='public'>
<ownedAttribute xmi:type='uml:Property' xmi:id='5'
name='textinput' visibility='private' aggregation='composite'
type='1_1_2_1'/>
<ownedAttribute xmi:type='uml:Property' xmi:id='6'
name='textinput' visibility='private' aggregation='composite'
type='1_1_2_2'/>
<ownedAttribute xmi:type='uml:Property' xmi:id='7'
name='button' visibility='private' aggregation='composite'
type='1_1_2_3'/>
</packagedElement>
------------------------------------------------------------------------------<packagedElement xmi:type='uml:Class' xmi:id='1_1_2_1'
name=''Username' ' visibility='public'/>
<packagedElement xmi:type='uml:Class' xmi:id='1_1_2_2'
name=''Password' visibility='public'/>
<packagedElement xmi:type='uml:Class' xmi:id='1_1_2_3'
name='Submit' visibility='public'/>
Figure 5. UML XML Description (Part1).
© 2012 ACADEMY PUBLISHER
138
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Fig. 5 is also divided into two parts as shown by the
dashed lines. In the first part, the block is defined as a
class type and the elements in this block are defined as a
property type. In the second part, the elements in block
are only defined as a class type.
This means that the group element, or in other words
UIContainer elements, is defined as a class type only, and
the primitive user interface (UIElement) element is
defined as a class and property type. Fig. 6 represents the
second part of Web UML XML description. In this part,
the coordinates for each element are specified by four
points as shown in bold text in Fig.7. These points are:
1. Padding-left: sets the left padding (space) of an
element.
2. Padding-Top: sets the top padding (space) of an
element.
3. Width: sets the width for element.
4. Height: sets the height for element.
These points are structured as: Padding-left, PaddingTop, Width, and Height.
<mdElement elementClass='Class' xmi:id='1004'>
<elementID xmi:idref='1_1_2'/>
<properties>
<mdElement elementClass='BooleanProperty'>
<propertyID>SUPPRESS_STRUCTURE</propertyID>
<propertyDescriptionID>SUPPRESS_STRUCTURE_DESCRIP
TION</propertyDescriptionID>
</mdElement>
</properties>
<geometry>90,225, 20, 130</geometry>
<propertyID>SUPPRESS_STRUCTURE</propertyID>
<compartment xmi:value='5^6^7'
compartmentID='ATTRIBUTES'/>
<parts>
<mdElement elementClass='Part' xmi:id='1005'>
<elementID xmi:idref='5'/>
<geometry>100,255, 20, 20</geometry>
<propertyID>SUPPRESS_STRUCTURE</propertyID>
</mdElement>
<mdElement elementClass='Part' xmi:id='1006'>
<elementID xmi:idref='6'/>
<geometry>100,280, 20, 20</geometry>
<propertyID>SUPPRESS_STRUCTURE</propertyID>
</mdElement>
<mdElement elementClass='Part' xmi:id='1007'>
<elementID xmi:idref='7'/>
<geometry>100,305, 20, 20</geometry>
<propertyID>SUPPRESS_STRUCTURE</propertyID>
</mdElement>
</parts>
</mdElement>
<UWE_PROFILE:PRESENTATIONGROUP XMI:ID='1018'
BASE_CLASS='1_1_2'/>
<UWE_PROFILE: TEXTINPUT XMI:ID='1019'
BASE_CLASS='1_1_2_1'/>
<UWE_PROFILE: TEXTINPUT XMI:ID='1020'
BASE_CLASS='1_1_2_2'/>
XMI:ID='1021'
<UWE_PROFILE:BUTTON
BASE_CLASS='1_1_2_3'/>
-----------------------------------------------------------------------------<UWE_Profile: textInput
xmi:id='1030' base_Property='5'/>
<UWE_Profile: textInput
xmi:id='1031' base_Property='6'/>
<UWE_Profile:button
xmi:id='1032' base_Property='7'/>
Figure 7. UML XML Description (Part3).
Figure 8. Group of HTML Elements.
description as output file. Substantially, the mapping
process consists of the following steps:
1. Firstly, Contrast Map table between HTML elements
and UWE stereotypes, as shown in Table III.
2. Read the VIPS XML Description for each block, this
step continues until the end of VIPS XML
Description file is reached.
3. The VIPS XML attributes for each element are
examined, and then useful information from these
attributes such as the type of element and coordinates
are extracted.
TABLE III.
VIPS XML ATTRIBUTES AND THEIR DESCRIPTION.
HTML Element
<div> </div>
<span></span>
<fieldset></fieldset>
<select>
<input type=” radio”/>
<input type=”checkbox”/>
<button></button>
<input type=" reset " />
<input type=" button " />
<input type=" submit " />
<input type=" file " />
Figure 6. UML XML Description (Part2).
<input type=" text " />
<input type=" password " />
< textarea></ textarea>
Fig. 7 represents the third part of Web UML XML
description. In this part, the stereotype of element is
specified by UWE profile. Also, Fig. 7 is divided into two
parts as shown by the dashed lines, First part specifies the
stereotype of element from class type, and the second part
also specifies the stereotype of element from property
type, as is shown by bold text in Fig. 7.
C. Mapping Process Stage
The mapping process is considered the main part of
our approach. This process accepts the VIPS XML
description as input file and produces Web UML XML
© 2012 ACADEMY PUBLISHER
<a href=””></a>
<p> </p>
4.
5.
UEW stereotype
Group Presentation
Selection
Button
File upload
Text input
Anchor
Text
<form > </form>
Form
<img src=””></img>
Image
The extracted information for all elements in block is
stored together in the data structure list for block.
Examine the data structure list for each block to
check whether the block has a collection of elements
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
139
that have the same type or not. For example, check
whether the blocks contain a collection of anchors or
not.
6. Mapping location step is considered the most
important step; this step needs to change the
coordinates for elements to become consistent within
UML presentation model.
7. Mapping from HTML element to UWE Stereotypes.
8. Now, the content of data structure list for each block
is Web UML XML Description. This content is
stored into Web UML XML file.
Fig. 9 shows the pseudo code for mapping process.
1: Input: VIPS XML description for blocks in the segmented Web
Page.
2: Output: UML XML description for blocks in the segmented
page.
3: Begin
4: Counter Å 1
// Counter for blocks.
5: total_blocks Å N // N is the no. of blocks in segmented page.
6: For counter to total_blocks step by 1
7: BlockÅ list_of_blocks [counter]
8: Counter_elemnent Å 1
// Counter for elements in
block
9: While Block contains elements with VIPS XML description
10: ElementÅ Block [Counter_elemnent]
11:VIPS_ Info Å Extract useful information from VIPS XML att.
of Element.
12: WebUMLXML_info Å Mapping _to_WebUMLXML (VIPS_
Info).
13: Block [Counter_elemnent] Å WebUMLXML_info
14: Counter_elemnent INCREMNET BY 1
15: End While
16: Store the information of block and elements into Web UML
XML file.
17: End for
Figure 9. Mapping Process Pseudo Code.
D. UML Presentation Model Generation Stage
After mapping process stage, the Web UML XML
description is used as input file for Magicdraw tool to
generate UML presentation model. The UML
presentation model in Fig. 10 is generated by parsing the
WebUML XML Description in Fig. 5, Fig.7 and Fig. 7
which form the simple HTML block as shown in Fig. 8.
IV. IMPLEMENTATION DETAILS
The process described in section 3 is implemented
using C#.Net Programming language (.NET Framework
3.5). This language provides the object oriented
programming which assists in programming process
through providing the libraries which provides many
features. In addition, we use the following APIs:
1. microsoft.mshtml.dll
2. MSXML3.dll
3. PageAnalyzer.dll
The PageAnalyzer.dll is considered the main API used
in our implementation; this dll file generates the VIPS
XML Description for any Web Page.
The main challenges which we faced during the
implementation process are:
Figure 10. Simple UML Presentation Model.
1.
2.
Mapping the coordinates of element and block in
VIPS XML Description to coordinates in Web UML
XML Description.
Take into account all the HTML elements.
V. CASE STUDY OF
THE PROPOSED APPROACH
In this section, we will illustrate the reverse
engineering process described earlier by applying it to a
specific website.
A. Page Segmentation
Fig. 11 shows the home page for goalzz web site , this
page is segmented using the VIPS algorithm into blocks
as shown in Fig. 12, some of these blocks are small and
others are large in size depending on the structure of the
page. The page layout of the segmented page is shown in
Fig. 13. In addition, Fig. 14 shows the DOM tree for the
segmented page.
B. Mapping Process
After the page segmentation process, the VIPS
algorithm assigns each block, sub-block, and element
with XML descriptions; these descriptions capture some
properties of each block and sub-block in the web page,
then the VIPS XML format is mapped and transformed
into the Web UML XML format as illustrated in Fig. 2.
Figure 11.Web Page before Segmentation Stage.
© 2012 ACADEMY PUBLISHER
140
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
C. UML Presentation Model Generation Stage
After the mapping and transformation process is
achieved, the presentation model can be imported and
manipulated in the Magicdraw modeling tool. Fig.15
shows the presentation model for goalzz home page.
Figure 14. DOM Tree for Segmented Page.
Figure 12.Web Page after Segmentation Stage.
Figure 15. UML Presentation Model for Goalzz Homepage.
Figure 13. Page Layout for Goalzz Homepage.
VI. SUMMARY AND FUTURE WORK
In this paper we present an approach of reverse
engineering web applications into UML web presentation
model. This issue is evolved from the more generic
reverse engineering process, concentrating on the
structure of the web page. We have presented an
approach to the identification of structure in a web page
and model this structure in UML presentation model. The
approach relies on a number of structured techniques
such as page segmentation.
Future work will concentrate on building a complete
framework which automatically builds the UML
presentation model for any given application. The process
of presenting the UML presentation model will be
automated and will apply content mining along with the
© 2012 ACADEMY PUBLISHER
segmentation technique to accurately identify different
blocks of the web page.
Also, we will work on handling Dynamic HTML pages
(DHTML). By DHTML pages we mean pages which
change its layout on the client side. So you will have a
unique URL with different page layout segmentation. For
example consider a faculty member website, page simply
list the faculty member publication, if the user clicks on
any publication the abstract of that publication will be
shown on the page without the need to connect back to
the server. Handling DHTML pages introduces two
issues that need to be considered in our future work. First,
we need to make sure that the segmentation process takes
this into consideration. Second, we need to check if UWE
notation is flexible and rich enough to capture and model
such a dynamic behavior of the web page.
ACKNOWLEDGMENT
This research was supported in part by German
Research Foundation (DFG), Higher Council for Science
and Technology (HCST) and Jordan University of
Science and Technology (JUST).
REFERENCES
[1] G. A. D. Lucca, A. R. Fasolino, and P. Tramontana,
"Reverse engineering web applications: the WARE
approach," Journal of Software Maintenance and
Evolution: Research and Practice, v.16, n.1-2, p.71-101,
January-April 2004 “doi:10.1002/smr.281”.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
[2] E. J. Chikofsky and J. H. II Cross, "Reverse engineering
and design recovery: a taxonomy," Software, IEEE, vol.7,
no.1, pp.13-17, January 1990 “doi:10.1109/52.43044”.
[3] D. Cai, S. Yu, J.-R. Wen, and W.-Y. Ma, VIPS: a visionbased page segmentation algorithm, Microsoft Technical
Report, MSR-TR-2003-79, November 2003.
[4] S. Tilley and S. Huang, "Evaluating the reverse engineering
capabilities of Web tools for understanding site content and
structure: a case study,"
Proceedings of the 23rd
International Conference on Software Engineering, pp. 514523, May 2001 “doi: 10.1109/ICSE.2001.919126”.
[5] S. Chung and Y.-S. Lee, "Reverse software engineering
with UML for Web site maintenance," Proceedings of the
First International Conference on Web Information
Systems Engineering, vol.2, pp.157-161, 2000 “doi:
10.1109/WISE.2000.882874”.
[6] R. Virgilio and R. Torlone, "A Structured Approach to
Data Reverse Engineering of Web Applications,"
Proceedings of the 9th International Conference on Web
Engineering, San Sebastián, Spain, June 2009 “
doi:10.1007/978-3-642-02818-2_7”.
[7] B. Djelloul, M. Mimoun, and B. S. Mohamed, "Ontology
based Web Application Reverse-Engineering Approach,"
INFOCOMP Journal of Computer Science, vol. 6, pp. 3746, March 2007.
[8] C. Kroiß and N. Koch, "The UWE Metamodel and Profile
– User Guide and Reference", Technical Report 0802,
Ludwig-Maximilians-Universität
München,
p.
35,
February 2008.
Natheer Khasawneh is an Assistant Professor in the
Department of Software Engineering at Jordan University of
Science and Technology since 2005. He received his BS in
Electrical Engineering from Jordan University of Science and
Technology in 1999. He received his Master and PhD degrees
in Computer Science and Computer Engineering from
University Akron, Akron, Ohio, USA in the years 2002 and
© 2012 ACADEMY PUBLISHER
141
2005 respectively. His current research interest is data mining,
biomedical signals analysis, software engineering and web
engineering.
Oduy Samarah is a Master student in the Department of
Computer Engineering at Jordan University of Science and
Technology since 2009. He received his BS in Computer
Engineering from Jordan University of Science and Technology
in 2009. His current research interest is in Wireless sensor
network, software engineering and web engineering.
Safwan Al-Omari is an assistant professor in the department of
Software Engineering at Jordan University of Science and
Technology. Dr. Safwan Al-Omari received his PhD degree in
Computer Science from Wayne State University in 2009. He
received his Master degree in Computer and Information
Science from the University of Michigan-Dearborn and
Bachelor degree in Computer Science from the University of
Jordan in 2003 and 1995, respectively. His current research is in
software engineering, service-oriented computing, and cloud
computing..
Stefan Conrad is a Professor in the Department of Computer
Science at Heinrich-Heine-University Duesseldorf, Germany.
He was an Associate Professor in the Department of Computer
Science at Ludwig-Maximilians- University in Munich,
Germany, from 2000 to 2002. From 1994 to 2000, he was an
Assistant Professor at the University of Magdeburg where he
finished his ‘Habilitation’ in 1997 with a thesis on federated
database systems. From 1998 to 1999, he was also a Visiting
Professor at the University of Linz, Austria. He received his
PhD in Computer Science in 1994 at Technical University of
Braunschweig, Germany. His current research interests include
database integration, knowledge discovery in databases, and
information retrieval. He is a (co)author of two books (in
German) and a large number of research papers.
142
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Adaptive FEC Technique for Multimedia
Applications Over the Internet
M. AL-Rousan and . A. Nawasrah
Jordan University of Science & Technology
College of Computer and Information Technology
Irbid, Jordan
alrousan@just.edu.jo
Abstract— Forward Error Correction (FEC) is a common
technique for transmitting multimedia streams over the
Internet. In this paper we propose a new approach of
adaptive FEC scheme for multimedia applications over the
Internet. This adaptive FEC will optimize the redundancy of
the generated codewords from a Reed-Solomon (RS)
encoder, in–order to save the bandwidth of the channel. The
adaptation of the FEC scheme is based on predefined
probability equations, which are derived from the data loss
rates related to the recovery rates at the clients. The server
uses the RTCP reports from clients and the probability
equations to approximate the final delivery ratio of the sent
packets to the client after applying the adaptive FEC. The
server uses the RTCP reports also to predict the next
network loss rate using curve fitting technique to generate
the optimized redundancy in-order to meet certain residual
error rates at the clients.
Index Terms— Forward Error Correction (FEC), ReedSolomon coder, network loss prediction, redundant
bandwidth optimization.
I.
INTRODUCTION
Streaming of multimedia over the Internet suffers
many difficulties, because of the already installed
equipments and protocols that support mainly data
applications. Recently, many multimedia applications
over the Internet infrastructure are taking place.
Applications such as Voice over IP (VoIP), Video on
Demand (VoD) and interactive gaming are utilizing the
installed switches, routers and backbone of the Internet.
Using the Internet infrastructure ensures multimedia
applications with low cost to end users.
Although the TCP session guarantees the delivery of
all the of the packets; it is not appropriate for on-line or
interactive multimedia streams, because the out-of-order
discard mechanism and NACK-retransmit technique
generates unsuitable jitter of the play-back multimedia
player at the client. The data of multimedia streams are
attached to time, which means that the arrived packet is
useful only if it arrived before the play-back time. So the
Forward Error Correction (FEC) technique has been used
for multimedia applications, which mainly depends on
sending redundant packets that might be used to recover
the lost packets [1-3].
A. Related Work
Most of the FEC research concentrated toward bit
error recovery, where the Reed-Solomon codec is
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.142-147
extensively used for storage devices such as the Compact
Disks (CD) [15]. Recent network installations suffer very
low Bit Error Rates (BER), as low as 10-9 [15] in fiber
networks. The main loss contributor in the Internet is due
to buffering of packets and discard mechanisms in
routers. Also multimedia applications suffer the out-oftime delivery of packets, due to the latency of arrived
packets. The packet level recovery includes more
complex computations due to the large arrays involved in
the generation of large blocks of data. For a ReedSolomon codec the computational complexity is still
relatively low at the server, but it is much complex at the
client side [7].
The FEC source techniques vary based on the
application, Nonnenmacher et al [12] suggested a hybrid
FEC and ARQ layer for time tolerance multicast
applications, where the FEC is applied as the 1st layer,
then the normal ARQ procedure will take place for lost
packets after the FEC operation, Chan et al [4] also target
the time tolerance video streams by introducing another
selectively delayed replication stream rather than the FEC
scheme, in order to achieve certain residual loss
requirements. The work of Parthasarathy et. al. [13]
presents another hybrid approach for high-quality video
over ATM networks; by joining FEC technique at the
sender side with simple passive error concealment at the
client side, which in turn enhances the packet recovery
even at high network loss rates. Yang et al [16]
introduced a FEC scheme adapted for video over Internet,
based on the importance of high spatial-temporal frame
packets, and it its effect on further depending packets, so
they send multiple redundancy levels based on the spatial
content of the packets. Change, Lin and Wu [5] studied
the FEC impact for CDMA 3-D systems over wireless
ATM networks, so they presents two levels of FEC for
header and payload packets, the header contains the
Discrete Cosine Transform (DCT) information so it
requires powerful FEC scheme to be reliably delivered,
and the payloads will be transmitted with lower FEC
protection. Song L., Yu M., Shaffer M. [15] present ideas
for hardware designing blocks of Reed-Solomon coders.
B.
Forward Error Correction
In traditional FEC, the server adds n-k redundant
(parity) packets to every k data packets, this yields n
packets to be transmitted. At the client side, if any k
packets from the n packets were received then the client
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
can recover all the lost packets without retransmission.
The amount of parities n-k is determined at the start of a
session, where the redundancy is calculated based on a
long-term average of network loss. The redundancy R is
defined as the amount of parities n-k to the block of
packets k as in equation 1 [12]
R =
n−k
k
(1)
The generation of the extra parities requires
mathematical codec. The codec must be reversible, so the
client can reconstruct the lost data out of the received
data. The Reed-Solomon codec is often used [6,7].
In this paper we adaptively optimize the parity
packets n-k, in order to save the redundant bandwidth
without degrading the quality of the displayed media. In
our approach, the source generates the maximum allowed
redundant parities n-k using Reed-Solomon encoder, but
it only sends r parity packets that are required to
overcome the expected network loss.
To overcome the above weakness, we propose the
BOAFEC approach where the source uses the long-term
average only to determine the maximum allowed
redundancy R. The BOAFEC predicts the current
network loss using simple three-points curve fitting
technique. The network loss prediction is used to
determine the amount of parity packets r to be transmitted
with the block of data packets k.
A. Probability Equations and Residual Loss Calculation
The BOAFEC uses the Reed-Solomon encoder to
generate the maximum allowed redundancy packets n-k,
but only sends r parities. That is because the ReedSolomon involves large arrays computations, and it is
computationally efficient to design only one coder with
fixed and maximum possible parities.
The BOAFEC predicts the network loss then uses the
probability equations to calculate the expected loss at the
client. And since packets suffer only two cases, whither
delivered or lost, the Binomial Distribution is hold [7].
Assuming the loss of a packet is the event of success, and
applying the Binomial distribution for a group of n
packets. The probability of l packets to be lost from n
packets, if the loss probability is πv will be:
⎛ n⎞
P(Loss= l) = ⎜⎜ ⎟⎟*(1−πv )n−1 *πvn
⎝l ⎠
Figure 1. FEC group packets with k=4, n-k = 2 and n = 6
Our work of adaptive FEC shows a considered
bandwidth saving over networks with low to medium loss
rates without affecting the quality of the multimedia
applications. That is, our scheme saves about 25% of the
redundant bandwidth, which leads to more clients can
subscribe to the same server, also the proposed scheme
responds to networks with high loss rates by saturating to
the maximum allowed redundancy, which corresponds to
a best effort mode, where the adaptive FEC cannot save
the bandwidth but can achieve the same quality of a
Reed-Solomon FEC.
II.
BANDWIDTH OPTIMIZED ADAPTIVE FEC
(BOAFEC)
The traditional FEC approach determines the
redundant parities based on the long-term average of
network loss, which is not suitable for multimedia
applications where the loss is instantaneous. On the
counterpart; generating on-the-spot parities is not
possible since the source does not know the current loss
rate at the client side, also the generation of adaptive
parities involves more computational complexity. Also
the long-term average of loss can miss lead the source to
send more parities than required, which results in wasting
the bandwidth.
© 2012 ACADEMY PUBLISHER
143
(2)
Since the group of n packets includes k data packets
and r parity packets, the Reed-Solomon coder can recover
up to r different packets from the lost packets. So the
FEC coder has the property that if l packets were lost
from the above group, then two cases apply:
l ≤ r: All packets will be recovered
l > r: None of the packets can be recovered
The FEC function of order r is defined as:
⎧0, l ≤ r
FECr (l) = ⎨
⎩l, l > r
(3)
The Expectation function [8] of the number of lost
packets is:
⎛ n⎞
l
E(x) = ∑l⎜⎜ ⎟⎟*(1−πv )n−l *πv
1 ⎝l ⎠
n
(4)
And hence the Expectation function after applying the
FEC is:
⎛ n⎞
l
E(x) = ∑l⎜⎜ ⎟⎟*(1−πv )n−l *πv
r+1 ⎝ l ⎠
n
(5)
The Expectation function presents the number of
expected packets over a group of n packets, so new loss
144
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
rate πv‘ which is the same as the residual loss rate ξr
after applying the FEC operation is:
πv' =ξr =
E
n
•
•
•
(7)
And so the delivery rate D is given as:
D = 1- π’v
(8)
The above equations lead the source to calculate the
residual loss rate ξr. The source will adaptively increase
or decrease the redundant packets r, in order to meet the
specified residual loss ξs at the client.
The number of packets k and number of maximum
added codewords n-k will influent the Reed-Solomon
encoder, the largest block size will result in more
computational complexity. The L. Rizzo Reed-Solomon
codec [7] is commonly used for software FEC coding.
B. Network Loss Prediction
The network loss heavily fluctuates over time, so the
average of long periods rarely presents the actual network
status, Figure 2 shows the loss of a network running for
120 seconds, although Figure 2 presents the average of
500ms for each reading, it is still fluctuating over time.
The long-term average loss of this run was 15.3%, while
it is obviously deviate most of time on this average.
We present a simple network loss prediction based on
the last three RTCP reports from the destination to the
source. Every RTCP report contains the loss at the client
at last transmission. The source constructs a matrix of the
network loss, and using Gaussian Elimination method
and Pivoting, the source can predict the next loss rate in a
finite time. The source will then use the predicted loss to
send the appropriate number of parity packets r, which
can be used at the client to reconstruct the lost packets in
order to meet the specified residual loss rate ξs.
Start the transmission assuming the highest
network loss, and so it uses the maximum
redundancy.
Wait for three RTCP reports in order to predict
the next network loss rate.
Calculate the optimal redundancy based on the
probability equations then update the number of
redundant packet for the next transmission.
Client:
•
•
•
•
Ask for a reservation for a suitable bandwidth
for the Multimedia application.
Determine the acceptable residual loss rate ξs for
the application.
Specify a client window size based on the
round-trip-time, the block of packets k and the
Rmax
Send RTCP reports
D.
The BOAFEC Parameter
The BOAFEC generates the redundant parities based
on four parameters that are the Network Loss Rate
(NLR), the maximum available redundant bandwidth, the
maximum allowed jitter and the target residual loss rate
ξs.
Network Loss Rate (NLR):
The NLR apparently is the main influence factor of
the BOAFEC decision of the number of redundant
parities. The BOAFEC cannot have control over the
NLR, because the NLR corresponding to the network
failures or congestions. Also the NLR is the only
continuously changing variable of the four BOAFEC
parameters, because the other three parameters are fixed
at the start of a session.
C.
BOAFEC Procedure
The server and client negotiate at the start of the
session to determine the network status, such as the
round-trip-time and the long-term loss average. Also they
negotiate to determine the parameters of the operation
such as residual loss rate, number of packets k in each
block and the maximum allowed redundancy. The
procedure of operation for the BOAFEC is as follows:
Server:
•
•
Allocate a certain bandwidth for each client at
the start of a session based on it available
resources and the multimedia application.
Determine the k and n-k parameters, and hence
the maximum allowable redundancy from
equation (1) based on the long-term history of
the network loss and the target residual loss ξs.
Rmax =
n−k
k
© 2012 ACADEMY PUBLISHER
Figure 2: Raw Network Loss based on 120 seconds
simulated network, every reading represents an
average of 500ms.
Maximum Redundant Bandwidth (MRB):
The MRB determines how much redundancy can the
server afford for the client as redundant codewords to be
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
used later for recovering lost packets. The MRB inversely
proportional to the residual loss rate. The MRB is defined
as the number of redundant parity packets to the number
of data packets, equation 1 represent the MRB.
The server determines the MRB based on the
registered QoS for each client, the better QoS
requirement needs more MRB to be associated to the data
transmission, in –order for the receiver to reconstruct
more lost packets.
Maximum Allowed Jitter (MAJ):
The MAJ represent the maximum jitter for a packet to
be displayed in order for the media application to run
smoothly. The MAJ indicates the scale of interactivity for
the media application. Applications with higher inactivity
requires less MAJ. The MAJ leads the media player to
determine when the player should discard a delayed
packet.
Target Residual Loss Rate (ξs):
The target residual loss rate ξs is the maximum
tolerable loss rate for a multimedia application in order to
run smoothly. Since the multimedia applications cannot
tolerate the excess delay generated from the
retransmission of the lost packets, there must be a
residual loss even when using the FEC. The residual loss
can further be reduced by using receiver based error
correction techniques like the Interleaving or repetition of
lost packets [14].
III. NETWORK LOSS BEHAVIOR
The network loss behavior over the Internet is very
complex to be defined, because of the many variables that
cannot be predicted nor defined. Although the loss over a
channel is simply figured by the number of lost packets
related to the total number of sent packets, but the
sequence or probability of a packet to be lost is not
simply defined.
A. Gilbert Model
A well known approximation of network loss is the
Gilbert-Model, which uses Two-State Markov chain to
represent the end-to-end loss. The Gilbert model is
widely used to simulate the Internet loss, due to its
simplicity and mathematical traceability [9][12][16]. The
Two-State Markov is shown in Figure 3. The 0 state
represents a packet was lost, where the 1 state represents
a packet reached the destination.
Let p denote the transition from the 0 state to state 1,
and q denote the transition from state 1 to state 0, so the
probability of losing a packet is (1-q), and the probability
to lose n consecutive packets equals (1-q)qn-1. From the
Markov Chain transition matrix [16], the long run loss
probability π can be defined as:
π =
© 2012 ACADEMY PUBLISHER
p
p + q
(9)
145
Fig
ure 3: Two-State Markov Model
IV. SIMULATION RESULTS
In this section we present the simulation results for
the BOAFEC scheme, we show the response of the
BOAFEC with different MRB and how it effects the
overall residual loss ξs and the used redundancy. Also we
show the BOAFEC results and compare it with the pure
Reed Solomon FEC (RS FEC) and a replicated stream
approach.
The MRB is very important factor for improving the
recovery rates at the clients, whereas the server must
bound the MRB in order to determine the number of
clients that can be attached to it at once. The MRB relates
directly with recovery rates at the clients, and hence it
relates inversely with the residual loss. Using the
BOAFEC can let the server to assign more MRB to
clients since the BOAFEC scheme optimizes that
redundancy.
Figures 4 and 5 show the response of the BAOFEC to
different NLR when the MRB are varying. Figure 4
shows that increasing the MRB results in better residual
loss rates at the clients, but it also shows that increasing
the MRB up to a certain level for networks with low loss
rates, such as when the NLR = 0.05; where increasing the
MRB over the 0.2 results in slight reduction of the
residual loss at clients, this response was due to the
BOAFEC response when the residual loss matches the
target loss ξs. Figure 5 presents the relation of the MRB
to the real used redundancy, note that the BAOFEC
optimizes the redundant bandwidth so the parity packets
are used as much as possible.
We finally compare the BOAFEC with the traditional
Reed-Solomon FEC and with the replicated stream FEC
scheme. The replicated stream approach simply sends
every packet twice, this increases the probability that at
least one of the two packets could arrive. Apparently the
replication stream requires 100% of redundancy, with the
lowest processing overhead over the known FEC. The
replicated stream is very useful for devices with low
processing resources, also it is suitable for networks with
high loss rates.
The Reed-Solomon codec was simulated using the
parameters k = 30, n-k = 15 and hence n = 48. The
BOAFEC parameters was residual loss rate ξs = 1%, k =
30 and maximum redundancy = 0.6.
146
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Figure 4: Residual loss versus the MRB for different
NLR, k = 20 and ξs = 1%
Figure 5: The used redundant bandwidth versus the MRB
for different NLR, k = 20 and ξs = 1%
Figure 5: Redundant Bandwidth versus Network Loss. k
= 30, n-k = 18, ξs = 1% and maximum redundancy =
60%.
From Figure 4, the residual loss rate for the BOAFEC
scheme is very close to the RS FEC in the region of loss
of 0% to 20%, where both the BOAFEC and RS FEC
performs much better than the replicated stream in the
above region, but it is still known that the replication
stream performs better than both in networks with high
loss rates (greater than 40%). Figure 4 shows also a swing
results for the BOAFEC under the specified residual loss
ξs, which is 1% in our example. The residual loss ξs is
considered to be compensated, whither it is objectively
tolerable or the destination uses other client based repair
techniques, such as Interpolation or regeneration of lost
packets [14].
Figure 5 shows the redundant bandwidth required for
each scheme. Obviously the BOAFEC uses the lowest
bandwidth while maintain close residual loss rates. The
BOAFEC achieves its best results for networks with low
to medium loss rates. Note when the loss rate exceeds the
25%; the BOAFEC saturates to the limit of its maximum
allowed redundant bandwidth in order to reduce the
residual loss rate.
V. CONCLUSION
Figure 4: Residual loss versus the network loss. k = 30, nk = 18, ξs = 1% and maximum redundancy = 60%.
© 2012 ACADEMY PUBLISHER
The packets loss is inevitable in networks, data
networks can tolerate the latency but not the loss, where
multimedia networks can tolerate the loss but cannot
tolerate the latency, due to the interactive nature of
multimedia applications. The FEC presents the least
latency recovery technique. The FEC is a very promising
technique for developing Multimedia applications over
the Internet without scarifying the QoS of the media
applications.
In this paper, we proposed and study a bandwidth
optimized FEC approach, by predicting the loss at the
client, while optimizing the amount of redundancy in
order to achieve a certain residual loss rate.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
The BOAFEC achieves very close recovery rates to
the pure FEC, while saving 25% (on average) of the
bandwidth, when the network loss rates are in the range
of 0% to 20%. In networks with high loss rates, the
BOAFEC saturates on the maximum allowed redundancy
in order to achieve the best possible quality.
REFERENCES
[1] Chih-Heng Ke, Rung-Shiang Cheng2, Chen-Da Tsai, and
Ming-Fong Tsai, "Bandwidth Aggregation with Path
InterleavingForward Error Correction Mechanism for
Delay-Sensitive Video Streaming in Wireless Multipath
Environments", Tamkang Journal of Science and
Engineering, Vol. 13, No. 1, pp. 1-9 (2010).
[2] Tsai, M.-F., Shieh, C.-K., Hwang, W.-S. and Deng, D.-J.,
“An AdaptiveMulti-Hop FEC Protection Scheme for
Enhancing the QoS of Video Streaming Transmission
overWireless Mesh Networks,” International Journal of
Communication Systems, Vol. 22, pp.1297−1318 (2009).
[3] Edward Au, Francesco Caggioni, and Jeff, Hutchins, "Core
Optics100G Forward Error Correctio", White Paper, 2010,
www.oifforum.com, accessed June 21, 2011.
[4] Chan S., Zheng X., Zhang Q., Zhu W., Zhang Y., “Video
Loss Recovery with FEC and Stream Replication”, IEEE
Transactions on Multimedia, VOL. 8, NO. 2, April 2006.
[5] Chang P., Lin C., Wu M., “Design of Multimedia CDMA
for 3-D Stereoscopic Video over Wireless ATM
Networks”, IEEE Transactions on Vehicular Technology,
VOL. 49, NO. 2, March 2000.
[6] Lacan J., V. Roca, J. Peltotalo, S. Peltotalo, “ReedSolomon Forward Error Correction (FEC)”, draft-ietf-rmtbb-fec-rs-01.txt, (work in progress) , June 2006.
[7] Luby M., Vicisano L., Gemmell J., Rizzo L., Handley M.,
Crowcroft J. “The Use of Forward Error Correction (FEC)
in Reliable Multicast”, RFC 3453, December 2002
[8] Mathews J.H., “Numerical Methods for Computer Science,
Engineering and Mathematics”, 1st Edition, ISBN 0-13626565-0.
[9] Mizuochi T., “Recent Progress in Forward Error
Correction and Its Interplay With Transmission
Impairments”, IEEE Journals, VOL. 12, NO. 4,
July/August 2006.
[10] Moore A.W., “Probability Density in Data Mining”,
Carnegie Mellon University.
[11] Moore A.W., “Probabilistic and Bayesian Analytics”,
Carnegie Mellon University.
[12] Nonnenmacher J., Biersack E., Towsley D., “Parity-Based
Loss Recovery for Reliable Multicast Transmission”,
IEEE/ACM Transactions on Networking, VOL. 6, NO. 4,
August 1998.
[13] Parthasarathy V., Modistino J., Vastola K. S., “Reliable
Transmission of High-Quality Video over ATM
Networks”, IEEE Transactions on Image Processing, VOL.
8, NO. 3, March 1999.
[14] Perkins C., Hodson O., Hardman Y., “A Survey of Packet
Loss Recovery Techniques for Streaming Audio”,
University College London, IEEE Network, September
1998.
[15] Song L., Yu M., Shaffer M., “10 and 40 Gbs Forward
Error Correction Devices for Optical Communications”,
IEEE Journals of Solid State Circuits, VOL. 37, NO. 11,
November 2002.
[16] Yang X., Zhu C., Li Z., Lin X, Ling N., “An Unequal
Packet Loss Resilience Scheme for Video Over the
Internet”, IEEE Transactions on Multimedia, VOL. 7, NO.
4, August 20.
© 2012 ACADEMY PUBLISHER
147
Mohammad Al-Rousan is
currently A full professor at the
Department
of
Network
Engineering and Security, Jordan
University of Science and
Technology (JUST). He was
educated in KSA, Jordan and
USA. He received his BSc in
Computer Engineering from
King Saud University, Saudi
Arabia, in 1986. He received his
M.S. in Electrical and Computer
Engineering from University of Missouri-Columbia, MI, USA
in 1992. In 1996 , he was awarded the PhD in Electrical and
Computer Engineering from Brigham Young University, UT,
USA. He was then an assistant professor at JUST, Jordan. In
2002, he joined the Computer Engineering Department at
American University of Sharjah, UAE. Since 2008, he has been
the Dean of College of Computer and Information Technology
at JUST. He is the Director of the Artificial Inelegant and
Robotics Laboratory, and a co-founder for the Nano-bio
laboratory, at JUST. His search interests include wireless
networking, System protocols, intelligent systems, computer
applications, and Nanotechnology, Internet computing. Dr. AlRousan served on organising and program committees for many
prestigious international conferences. He is the recipient of
several prestigious awards and recognitions. He co-chaired
international conferences on Information and Communication
Systems (ICICS09 and ICICS11).
Ahmad Nawasrah Received his MS in Computer Engineering
from Jordan University of Science and Technology in 2007. He
has been the COE of AL-Bahith Co. His research interests
include embedded systems and coding.
148
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Semantic Approach for the Spatial Adaptation of
Multimedia Documents
Azze-Eddine Maredj and Nourredine Tonkin
Research Center on Scientific and Technical Information (CERIST)
Ben Aknoun, Algiers, Algeria
Email: {amaredj, ntonkin}@mail.cerist.dz
Abstract—With the proliferation of heterogeneous devices
(desktop computers, personal digital assistants, phones),
multimedia documents must be played under various
constraints (small screens, low bandwidth, etc). Taking
these constraints into account with current document
models is impossible. Hence, generic source documents must
be adapted or transformed into documents compatibles with
the target contexts. The adaptation consists of modifying
this specification of the document in a minimal way to lead
it to satisfy the target profile. The profile defines constraints
that must be satisfied by the document to be played. At this
level, the transgressive adaptation becomes necessary when
no specification model exists to satisfy this profile. We focus
on the spatial dimension of a multimedia document and we
provide an approach for the spatial adaptation that permits
to best preserve the initial document semantic. In the
proposed transgressive adaptation, the relations that do not
comply with the target profile are not most often replaced
by the closest ones due to the fact that their immediate
neighboring relations have the same similarity degree
whereas there may exist differences between them. In this
paper, we extend this approach to, firstly, best preserve the
proximity between the adapted and the original documents
by weighting the arcs of the conceptual neighborhood
graphs and secondly, to deal with complex relations models
by integrating the concept of the relations relaxation graphs
that permit to handle the distances defined within the
relations.
Index Terms—multimedia document, spatial adaptation,
conceptual neighborhood graph, distance relaxation
I. INTRODUCTION
Multimedia document should be able to be played on
different platforms (mobile phones, PDAs, laptops,
PCs…) and must be presented according to the user
preferences. The challenge is to execute and transmit to
users media objects with a high quality in a consistent
presentation that reproduces as closely as possible the
semantics of the original document. Indeed, in the
interactive online courses, the reader can use any type of
devices without worrying whether the document is
supported and the provided document shall reproduce the
same content as specified by the author.
To deal with the diversification of target contexts and
user preferences, multimedia document must to be
adapted before being played, i.e., from the profile
(hardware and software constraints and user preferences)
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.148-154
and the source document, the adaptation must transform
the document to be compatible with the target profiles.
A lot of works has been done in this way and showed
that there are two types of adaptation. The first type is to
specify directly the different organizations of the
document for each target platform. In this case, the task
becomes complex since it requires an extra work to the
author when he should specify the conditions of
execution of his document on each target context. It can
also be incomplete since the author must foresee all the
existing targets. The second type of adaptation is based
on the dynamic adaptation of the document performed by
a program transforming the document. In this type, two
kinds of adaptation are possible: a local adaptation that
considers the media object individually but does not most
often, preserve the document semantics and the global
adaptation related to the document composition
(temporal, spatial and hypermedia specifications) and
which preserves the semantics of the document [1].
This paper focuses on the latter type of the adaptation
and concerns only the spatial aspect. It is devoted to the
transgressive adaptation of the spatial relations where the
document is represented by an abstract structure
expressing all relations between media objects. In this
context, adapt a multimedia document consists in
transforming its abstract structure in a manner that it
meets the requirements of the target profile.
In [1], it was shown that a spatial relation is
represented by the combination of the Allen relations [2]
on the horizontal and vertical axes and the spatial
adaptation follows the same principle as the one defined
for the temporal adaptation in [3]. This principle says that
if no model of the original specification satisfies the
adaptation constraints (context constraints) then
transgressive adaptation is applied. In this approach, the
transgressive adaptation consists in transforming the
relations between media objects while ensuring two main
properties: (i) the adaptation constraints are satisfied and
(ii) the adapted document is, semantically, as close as
possible to the initial document. This consists in finding
another set of models (solutions), close to the initial,
which satisfies these constraints. The proposed solution is
to replace each relation that does not meet the profile by
another semantically close. To find the closest relations,
the conceptual neighborhood graph proposed in [4] is
used.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
In this approach, the specification is done using models
(temporal and spatial) where delays and distances
between media objects are not considered, when actually,
the produced documents (which are subjects of
adaptation) are often, for expressiveness purpose,
composed using very complex models.
In our previous works [5], we showed the interest of
the weighted conceptual neighborhood graph usage when
seeking for the closest substitution relations.
In this work, we will extend this approach to spatial
models where distances between media objects are
defined. To represent such spatial models, we use the
temporal model of Wahl and Rothermel [6] which is an
adaptation of the general model of Allen [2] where the
authors proposed a set of composition operators
integrating the delay concept in the relations. The
elaboration of those operators was motivated by the need
to facilitate the temporal specification of the document.
More details about this representation are presented in
section 4.
In the second section of this article, we present the
context of our work. The third section presents
multimedia adaptation approaches and in section 4, we
present our approach to the spatial adaptation. In the fifth
section we present the adaptation procedure and the last
section concludes this paper.
II. CONTEXT OF THE WORK
A multimedia document specification comprises
temporal relations defining inter media objects
synchronization and spatial relations that express the
spatial representation of these media objects. The
different users of multimedia documents impose different
presentation constraints on the specification like display
capabilities (screen size and resolution). The user’s
device may do not have the necessary capabilities to
support the spatial constraints of the document.
For example, let us consider a multimedia document
with the following spatial relation: image B Above image
A and the resolution of the two images is 200 x 300. If the
terminal resolution is superior or equal to 400 x 600 then,
the user will not have any issue when displaying this
document but, if its resolution is inferior to 400 x 600, we
can have the following solutions [6]: (i) delete one of the
two images or (ii) resize image A or image B or (iii)
change the spatial relation Above by another relation.
The deletion of one of the two images may alter or
produce an incomprehensible document. Resizing one of
the objects will not affect the relation between them but
may lead to a wrong interpretation or make the image
indistinguishable in the case of an X-ray radio for
example. The spatial relation modification does not cause
information waste as in the image deletion and does not
make indistinguishable the images. It only changes the
places of those images.
Here, we focus on the relations transformation while
trying to preserve the document semantic as well as
possible.
© 2012 ACADEMY PUBLISHER
149
Before we present the spatial adaptation of multimedia
documents, we start by giving the different approaches
for multimedia documents adaptation.
III. ADAPTATION APPROACHES
Several approaches have been proposed for the
multimedia documents adaptation and we group them
into four categories: the three categories, specification of
alternatives, using transformation rules and using flexible
documents models as presented in [1] to which we added
the semantic and dynamic approaches.
Specification of alternatives: The author of the
document specifies a set of presentation alternatives by
defining criteria on some media objects of the document.
If the media objects satisfy the criteria then, they are
selected and presented else, they are deleted from the
presentation. The adaptation is performed beforehand like
in SMIL which defines the operator switch to specify the
alternatives that are played only if they comply with the
target profile. The advantage of this family is that the
adaptation is instantly. However, the author has to foresee
all the possible target profiles and specify all the
conceivable alternatives.
Using Transformation rules: This category uses a set
of transformation rules that are applied to the multimedia
documents. The adaptation consists on selecting and
applying rules to transform the document to satisfy the
target profile. The advantage of this approach is that the
author has not to care about the execution context of his
document. Furthermore, this set of rules can be
completed if new contexts appear. However, the entire
transformation rules should be specified to ensure an
efficient adaptation.
Using flexible documents models: The adapted
document is generated automatically from a noncomposed set of media objects represented by a model
defining an abstraction of the document. Thanks to a
formatting model, a multimedia presentation may be
generated. In this category, we can mention Guypers [8]
that aims to generate web-based presentation for
multimedia databases. It is based on the use of semantic
relations between multimedia objects and ontologies.
Semantic and dynamic approaches: In [1], an approach
based on the qualitative specifications of the document
was proposed. Each document is considered as a set of
potential executions and each profile is considered as a
set of possible executions. The adaptation is done
according to the context at the execution time. It consists
on calculating the intersection between the potential
(initial specification models) and the possible executions
corresponding to the target profile. The advantage of this
approach is its independence from description languages.
However, the usage of the conceptual neighborhood
graphs where all the weights of the arcs are set to one
(01) assumes that a relation may be replaced by any one
of its immediate neighbors while there are substantial
differences between them; especially when using more
elaborated relations models. Furthermore, the delays and
the distances defined within the relations are not
considered.
150
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
In the reminder of this paper, we present our
proposition for the spatial adaptation which fits into this
last category.
IV.
SPATIAL ADAPTATION
A. Context Description [9]
To describe the target context (profile), the universal
profiling schema (UPS) [10] can be used. It is defined to
serve as a universal model that provides a detailed
description framework for different contexts. UPS is built
on top of CC/PP and RDF [11]. Unlike CC/PP, UPS does
not describe only the context of the client but it describes
also all the entities that exist in the client environment
and that can play a role in the adaptation chain from the
content server to the target client.
UPS includes six schemas: the client profile schema,
the client resources profile, the document instance profile,
the adaptation method profile and the network profile.
The description of the user is included in the client profile
schema.
B. Spatial Relations Model
To describe the spatial presentation of a multimedia
document, we use the directional representation [12] that
permits to define the orientation in space between media
objects.
In this representation, a media object is considered as
two intervals corresponding to its projection on the
horizontal and vertical axes. The set of the directional
relations is obtained by combining the intervals of the
two media objects on the two axes by using the Wahl and
Rothermel relations model [6] presented in figure 1, on
each axis.
There are 20 possible relations (10 basic relations and
theirs inverses) on each axis. Thus, a spatial relation is
represented by two temporal relations [1]: one on the
horizontal axis and one on the vertical axis. For example,
the spatial relation left_top can be represented by its two
components: before (on the horizontal axis) and before
(on the vertical axis).
In the multimedia document spatial adaptation, we
follow the same procedure as the one presented in [5] for
the temporal adaptation.
C. Conceptual Neighborhood Graph of the Spatial
Relations
The representation of the spatial relations by two
components (temporal relations) permits us to use the
conceptual neighborhood graph of the temporal relations
[4] to elaborate the spatial relations conceptual
neighborhood graph. Indeed, for each component of the
spatial relations (vertical and horizontal components), we
use the temporal relations conceptual neighbourhood
graph. The composition of the two graphs gives the
conceptual neighborhood graph of the spatial relations.
It’s the square product of the conceptual neighborhood
graph of temporal relations.
Conceptual Neighborhood
Two relations between two media objects are
conceptual neighbors if they can be directly transformed
into one another by continuous deformation (shortening
or lengthening) of the duration of the media objects
without going through an intermediate relation.
For example, in figure 2, the relations before and
Overlaps are conceptual neighbors since a temporal
extension of the media object A may cause a direct
transition from the relation before to the relation
Overlaps. And in figure 3, the relations before and
Contains are not conceptual neighbors, since a transition
between those relations must go through one of the
relations Overlaps, Endin, Cobegin, Coend, Beforeendof1
, Cross-1, Delayed-1 or Startin-1.
The Conceptual neighborhood graph, presented in
figure 4, is defined as a graph where the nodes
correspond to the relations of Wahl and Rothermel model
and each arc between two nodes (relations) r and r’
corresponds to the satisfaction of the propriety of the
conceptual neighborhood, i.e., r and r’ are conceptual
neighbors.
Figure 2. Example of neighboring relations.
Figure 1. The relations model of Wahl and Rothermel [6]
© 2012 ACADEMY PUBLISHER
Figure 3. Example of non-neighboring relations
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
151
Figure 6. Information items of a relation
TABLE I.
INFORMATION ITEMS THAT CHARACTERIZE A TEMPORAL RELATION
Information
1
2
3
4
5
6
Signification begin(A) begin(B) end(A) end(B) 1>2 1<2
Information
7
8
9
10
11 12
1>4
1<4
3>4
3<4 3>2 3<2
Signification
Figure 4. Conceptual neighborhood graph of the Wahl and Rothermel
relations
Figure 5. Conceptual neighborhood graph of Allen relations
Weighting of the Conceptual Neighborhood Graph
In the conceptual neighborhood graph of the relations
of Allen [2] (figure 5) presented in [3], the weights of the
arcs are set to 1. This assumes that a relation can be
indifferently replaced by any one of its neighbors having
the same smallest conceptual distance whereas there may
be a substantial difference between the candidate
relations. It would be interesting to differentiate the
proximity degree between these relations. The distinction
in the proximity of the neighboring relations is done by
assigning different weights to the arcs of a graph.To
assign different weights to the arcs of a conceptual
neighborhood graph, our idea is to identify all the
information items that characterize a temporal relation so
they serve as a basis for comparing and differentiate the
similarity between the relations.
Information Items of a Relation
The analysis of a relation between two media A and B
(Figure 6) on a time axis showed that the positioning is
done according to the order that exists between their
respective edges (occurrence order of the beginning and
ending instants of the media objects). Therefore, to
characterize a temporal relation, we selected the
following information items: the values of the beginnings
and the endings of the media objects and the orders
(precedes (>) or succeeds (<)) between their edges.
Table1 gives a recapitulation of the selected
information items. Then, we have for each relation,
determined the information items that it contains among
the selected ones and the result is given in Table 2.
© 2012 ACADEMY PUBLISHER
TABLE II.
INFORMATION ITEMS CONTAINED IN THE WAHL AND ROTHERMEL
RELATIONS
Before
Overlaps
Endin
Cobegin
Coend
Beforeendof-1
Cross-1
Delayed-1
Startin-1
While
Contains
Beforeendof
Cross
Delayed
Startin
Cobegin-1
Endin-1
Coend-1
Overlaps-1
Before-1
1
0
1
0
1
0
0
1
1
1
1
1
1
1
1
1
1
1
0
1
1
2
1
1
1
1
0
1
1
1
1
1
1
0
1
1
1
1
0
0
1
0
3
1
1
1
0
1
1
1
1
1
1
1
0
1
1
0
0
1
1
1
0
4
0
1
1
0
1
0
1
1
0
1
1
1
1
1
1
0
1
1
1
1
5
0
0
0
0
0
0
0
0
0
1
0
1
1
1
1
1
1
1
1
1
6
1
1
1
1
1
1
1
1
1
0
1
0
0
0
0
0
0
0
0
0
7
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
8
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
9
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
10
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
11
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
12
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Calculation of the Similarity Degree Between a Relation
and its Immediate Neighbors
To calculate the similarity degree between a relation
and its neighbors, a distance should be defined. The aims
of this distance is to make the difference of proximity
between the relations but not to have a precision. Thus,
any distance definition can be used (Euclidian distance,
Manhattan distance, etc). We choose the Manhattan
distance defined as follows:
Let us consider the two vectors V (v1, v2... vn) and U
(u1, u2... un). The Manhattan distance between V and U is:
d (V-U) = Σni=1 |vi - ui|
In our work, we consider the information items of each
relation as a vector where the value of each information
item is set to 1 whether this information item is included
in the relation or zero (0) otherwise.
Using the Manhattan distance, we established the
distances between each relation and its immediate
neighbors as presented in table 3. Then, we affect the
calculated distances to the arcs of the conceptual
neighborhood graph as shown in figure 7.
152
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
TABLE III.
DISTANCES BETWEEN THE RELATIONS AND THEIRS NEIGHBORS
opportunity to best preserve the semantic of the initial
document.
For example, the relation A Before (d) B can be
transformed into the relation A Before (-) B where the
symbol "-" means that the distance is not specified.
We present below, the delays relaxation of the
relations of the Wahl and Rothermel model. For this, we
use a graph structure (we call it relaxation graph) based
on the number of delays defined for a relation as shown
in figure 8. The relaxation graph of a relation is
composed by the different forms of this relation obtained
by a progressive relaxation of its delays.
To illustrate this principle, consider the example of the
figure 9 where the specified relation is
Image <Overlaps (200,50,250>, Contains (10, 90)> Text
And the target screen size is 480 x 360.
The two objects cannot be displayed because the width
of the area they occupy (500 px) is larger than the screen
width (480 px). The distances relaxation of this relation
can lead to an adaptation solution without changing the
relation and this solution (see figure 10) may be:
Image <Overlaps (180, 70, 230>, Contains (10, 90)> Text.
Figure 7. Weighted conceptual neighborhood graph of Wahl and
Rothermel relations
The Conceptual Distance of a Spatial Relation
The conceptual distance [1] between two relations r
and r' is the length of the shortest path between those
relation in the conceptual neighborhood graph.
The conceptual distance of a spatial relation is defined
as the sum of the conceptual distances of its two
components (temporal relations on the two axes). For
example, let’s consider the two spatial relations
r1=<Before, Overlaps> and r2=<Endin, While>. The
conceptual distance between r1 and r2 is dx (r1, r2) + dy (r1,
r2) i.e, d (Before, Endin) + d (Overlaps, While).
C. Distance Relaxation
According to the consideration of a spatial relation as
two temporal relations, its distances are taken as delays in
the two temporal relations. In this case, before seeking for
substitution relation, we start by trying to relax these
delays in order to keep the relations and give us the
© 2012 ACADEMY PUBLISHER
Figure 8. Delays relaxation graphs
Figure 9. Example of a spatial specification
Figure 10. An adaptation solution with distance relaxation
Figure 11. An adaptation solution without distance relaxation
If we proceeded directly by replacing the relation as
was proposed in [1] then, the relation Overlaps will be
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
directly replaced by the relation While and we will get the
result of the figure 11.
V. ADAPTATION PROCEDURE
The semantic adaptation of a multimedia document is
achieved by modifying the specification of the document.
This involves finding another set of values for the
distances of the relations or otherwise a set of relations
satisfying the adaptation constraints of the target platform
with the smallest distance from the initial specification.
The distances relaxation and the relations replacement
processes are done by traversing the graphs (relaxation
graph for distances relaxation and conceptual
neighborhood graph for the relations replacement) in both
directions. This is done by searching the shortest path
between the relation to be replaced and the other relations
of the graph. A relation is considered as candidate to
replace the initial only if it does not lead to an
inconsistency (for this we recommend the use of the
Cassowary resolver [13]). Finally, the solution with the
smallest conceptual distance is considered as the adapted
document.
The adaptation procedure is done by following two
phases: Distances relaxation and transgressive adaptation.
A. Adaptation Algorithm
We implemented this procedure throw an algorithm as
shown in algorithm 1.
// Phase 1: relaxation
Input : MI1j: //Matrix of the document relations
// relaxation relations search
for i = 0 to n-1 do // n number of objects
for j = 0 to n-1 do
RG = SelectRG(MI[i, j]);
MS[i, j] = determineRelaxedRelations(RG);
end for
end for
// Elaboration of the possible combinations
//output : combinations list Cp
Cp = ElaborateCombinationsMatrix(MSij);
// Sort combinations according to the conceptual distance
for i=0 to nCombinations -1 do
d[i] Å 0;
// Matrix of the conceptual distances
for j=0 to n-1 do
d[i] = d[i] + Djikstra(C[i,j], MR[i,j]);
end for
end for
QuickSortCombinations(C[i], d[i]);
// Consistency Verification
found Å false ;
for i = 0 to nCombinations -1 do
if Consistency (C[i]) then
Solution Å (C[i]);
Exit();
end if
end for
// Phase 2: replacement
// replacement relations search
Clear (MSi, j);
for i = 0 to n-1 do // n number of objects
for j = 0 to n-1 do
for k=1 to NR do // NR : number of the relations of the model
© 2012 ACADEMY PUBLISHER
153
if respecteProfile (Rm [k]) then //Rm set of the model
relations
MS [i, j] Å MS [i, j] ∪ {Rm [k]};
end if
end for
end for
end for
// Elaboration of the possible combinations
//output : combinations list Cp
Cp = ElaborateCombinationsMatrix(MSij);
// Sort combinations according to the conceptual distance
for i=0 to nCombinations -1 do
d[i] Å 0;
// Matrix of the conceptual distances
for j=0 to n-1 do
d[i] = d[i] + Djikstra(C[i,j], MR[i,j]);
end for
end for
QuickSortCombinations(C[i], d[i]);
// Consistency Verification
found Å false ;
for i = 0 to nCombinations -1 do
if Consistency (C[i]) then
Solution Å (C[i]);
Exit();
end if
end for
Algorithm 1. Adaptation algorithm
- RG : relaxation graph of the selected relation MI[i, j]
- determineRelaxedRelations(RG) : determines all the
relations forms obtained by relaxing the initial relation by
using the correspondent Relaxation graph GR.
- ElaborateCombinationsMatrix(MSij): determines all
the possible combinations of the candidate relations for
the replacement of the initial relations of the document.
- Consistency (C[i]): Calls the linear constraints solver:
Cassowary for the consistency verification of the
solution.
B. Algorithm Description
Distance Relaxation
The algorithm takes as input the matrix MIij: matrix of
the complete relations graph of the initial specification.
For each relation of MIij, we identify its corresponding
RG from which we determine all the different form of
this relation obtained by the relaxation process (only the
relations that comply with the target profile are retained).
The result is placed in the substitution matrix MSij. Then,
we determine, by combinations, all the possible solutions
Ci from the matrix MSij. Next, we perform an ascending
sort of all solutions of Ci with the classical sorting
algorithm "quick sort" using the conceptual distances
calculated by using the Dijkstra's shortest path algorithm
in the GR of each relation. This will ensure that the
solutions are sorted from the closest specification to the
farthest from the original. Finally, we call the constraints
solver (cassowary [13]) for the consistency verification
and the calculation of the solution (new values to the
distances) for each specification in the order defined by
the sort. The first meet with a consistent solution stops
the process and that solution is considered as the adapted
document.
154
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Relations replacement
If at the end of the distances relaxation phase, no
solution was found, we perform the replacement of the
relations:
We take back the initial relation matrix MIij and we
determine the substitution matrix MSij, which gives for
each relation of the matrix Mij the relations candidates for
its substitution (not its relaxation as in the first phase)
from those that meet the target profile constraints among
the relations of the model, we determine by
combinations, all the possible solutions Ci from the
matrix MSij. Next, we perform an ascending sort of all
solutions of Ci using the conceptual distances calculated
by using the Dijkstra's shortest path algorithm through
the conceptual neighborhood graph. Finally, we call the
constraints solver cassowary for the consistency
verification and the calculation of the solution for each
specification. The first verification that gives a
consistency stops the process of solutions consistency
verification.
VI. CONCLUSION
In this paper, we proposed an approach to the spatial
adaptation of multimedia documents that takes into
account the distances defined in the relations.
The differentiation in similarity degrees between each
relation and its neighbors by affecting different weights
to the arcs of the conceptual neighborhood and the
introduction of the relaxation principle to take into
consideration the distances permit to replace a relation
with its closest semantically and thus to keep the
document as close as possible to the original document.
The first axe of our future work would be to merge the
two phases of the adaptation procedure by integrating the
RGs of the relations in the conceptual neighborhood
graph and then, have only one phase that will permit to
have an adaptation solution composed by relaxed and
replaced relations.
The second axe would be to determine the similitude
measure between the adapted document and the original
one by using some extra information (annotations) like
weights assigned to relations based on their importance in
the specification to determine relations to be modified or
to be removed if it’s necessary.
REFERENCES
[1]. S. Laborie. Adaptation sémantique de document
multimédia. PhD Thesis, University of Joseph Fourier Grenoble 1, France, 2008.
[2]. J. F. Allen. Maintaining knowledge about temporal
intervals. Communication of the ACM, vol. 26 (1), 1983,
pp. 832–843
[3]. J. Euzenat, N. Layaïda, V. Diaz. A semantic framework for
multimedia document adaptation. 18th International Joint
Conference on Artificial Intelligence (IJCAI). Acapulco
(MX), 2003, pp. 31-36.
[4]. C. Freksa. Conceptual neighborhood and its role in
temporal and spatial reasoning. In Proceedings of the
IMACS International Workshop on Decision Support
Systems and Qualitative Reasoning. Toulouse, France,
1991, pp. 181-187.
© 2012 ACADEMY PUBLISHER
[5]. A. Maredj, N. Tonkin. Weighting of the conceptual
neighborhood graph for multimedia documents temporal
adaptation. Proc. Of the International Arab Conference on
Information technology (ACIT’2010). Benghazi-Libya,
2010, pp 98.
[6]. T. Wahl, K. Rothermel. Representing Time in Multimedia
Systems. In Proceedings of the International Conference
on Multimedia Computing and Systems. Boston,
Massachusetts. 1994, pp. 538-543.
[7]. A. Gomaa, N. Adam, V. Atluri. Adapting Spatial
Constraints of Composite Multimedia Objects to achieve
Universal Access. IEEE International Workshop on
Multimedia Systems and Networking (WMSN05), held in
conjunction with the 24th IEEE International
Performance,
Computing,
and
Communications
Conference (IPCCC). Phoenix, Arizona, USA, April 2005.
[8]. J. Geurts, J. V. Ossenbruggen and L. Hardman:
Application specific constraints for multimedia
presentation generation. In Proceedings of the
International Conference on Multimedia Modeling, 2001,
pp. 247-266.
[9]. T. Lemlouma , N. Layaïda context-aware adaptation for
mobile devices (2004) in proceedings of the IEEE
international conference on mobile data management.
Berkeley, CA, USA, 2004, pp.19-22,
[10]. . Lemlouma , N. Layaïda, Universal Profiling for Content
Negotiation
and
Adaptation
in
Heterogeneous
Environments, W3C Workshop on Delivery Context, W3C
INRIA Sophia-Antipolis, France, 4-5 March 2002.
[11]. Graham Klyne et al, Composite Capability/Preference
Profiles
(CC/PP):
Structure
and
Vocabularies,
http://www.w3.org/TR/CCPP-structvocab/, W3C Working
Draft, 15 March 2001
[12]. D. Papadias, T. Sellis. Qualitative Representation of
Spatial Knowledge in Two-Dimensional Space. Special
Issue on Spatial Databases, Vol. 3(4), 1994, pp. 479-516.
[13]. J. Badros, A. Borning, S. J. Stuckey. The Cassowary linear
arithmetic constraint solving algorithm. ACM Transactions
on Computer-Human Interaction (TOCHI), vol. 8(4), 2001,
pp. 267–306.
Azze-Eddine Maredj acquired his
doctorate in computer science at the
university of science and technology
Houari Boumediane (Algeria). He is a
search scientist at the Research Center
on Scientific and Technical Information
(CERIST), algiers (Algeria) where he
lead the “multimedia system and
structured documents” team. His Research interests are
mutlimedia authoring systems and retrieval systems of
structured and multimedia documents. Currently, he works on
project devoted to delever an authoring system which permits
the edition, the rendering and the exploitation (indexation and
retreival) of mutlimedia documents.
Nourredine Tonkin received the
engineer degree in computer science,
from the Mouloud Mammeri University,
Tizi-Ouzou (Algeria) in 2004. He Works
at the Research Center on Scientific and
Technical
Information
(CERIST),
Algeria. His Research interests are
mutlimedia
authoring
systems.
Currently, he works on the adaptation of
multimedia documents.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
155
Adaptive Backoff Algorithm for Wireless Internet
Muneer O. Bani Yassein
Jordan University of Science and Technology /Department of Computer Science, Irbid, Jordan
Email: masadeh@just.edu.jo
Saher S. Manaseer and Ahmad A. Momani
University of Jordan/Department of Computer Science, Amman, Jordan
Jordan University of Science and Technology /Department of Computer Science, Irbid, Jordan
Email: saher@ju.edu.jo, ahmed_cs2004@yahoo.com
Abstract— The standard IEEE 802.11 MAC protocol uses
the Binary Exponential Backoff algorithm. The Binary
Exponential Backoff makes exponential increments to
contention window sizes. This work has studied the effect of
choosing a combination between linear, exponential and
logarithmic increments to contention windows. Results have
shown that choosing the right increment based on network
status enhances the data delivery ratio up to 37% compared
to the Binary Exponential Backoff, and up to 39 %
compared to the Pessimistic Linear Exponential Backoff
algorithms for wireless Internet.
limited transmission range; therefore each node in the
network connects to the nearest base station within its
transmission range [23].
Index Terms— Wireless Internet, MAC, CW, Backoff
algorithms
I. INTRODUCTION
The first appearance of wireless networks was in
1970s. Since that time, these networks are being
developed so fast [22, 23]. It has been noticed that in the
last decade all trends moved toward wireless Internet
technology [23]. Also the mobile wireless network which
is also called mobile ad hoc network has become the new
age of wireless networks. We can distinguish two types
of networks; infrastructure and ad hoc networks [1, 3, 22,
23]. See figures 1 and 2.
Figure 1: An example of an Infrastructure Wireless Internet [29]
Fig. 1 shows a simple example of the first type of
infrastructure wireless networks. Communication
between nodes at such networks is managed via a base
station or a central access point. Each base station has a
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.155-163
Figure 2: An example of Wireless
Internet
On the other hand, Fig. 2 shows an example of the
second type of wireless internet, a mobile ad hoc network
(MANET). A MANET is a set of mobile nodes that
communicate through wireless links. Each node acts as a
host and router. Since nodes are mobile, the network
topology can change rapidly and unpredictably over time
[1, 3]. In other words, a MANET does not have a base
station, so communication between nodes is managed by
the nodes themselves. Moreover, nodes are not expected
to be fully-connected, hence nodes in a MANET must use
multihop path for data transfer when needed [24].
Recently, most interests were focused on MANETs
due to potential applications provided by this type of
networks such as military operation, disaster recovery,
and temporary conference meetings.
A. Features and Characteristics of MANETs
Despite a MANET has many features shared with
infrastructure networks, it also has its own additional
features. Some of these features are:
• Dynamic network topology: nodes in the network are
free to move unpredictably over time. Thus, the
network topology may change rapidly and
unpredictably. This change may lead to some serious
issues, such as increasing the number of transmitted
messages between nodes of the network to keep
156
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
routing information table updated, and this will
increase the network overhead [25].
•
Distributed operations: in MANET there is no
centralized management to control the network
operations like security and routing, therefore, the
nodes must collaborate to implement such functions.
In other words, the control management is distributed
among nodes of the network [22].
• Limited resources: in MANETs nodes are
mobile, so they suffer constrained resources
compared to wired networks. For example,
nodes in a MANET depend on batteries for
communication and computation, so we
should take in to account how to optimize
energy consumption [26, 27, 28].
B. Applications of MANETs
MANETs are deployed in different environments due
to its valuable features of mobility, no base stations,
Some of its applications are [22, 23]:
• Military Operations
In battlefield environments, a MANET can be very
useful to setup a reliable communication between
vehicles and solders where it seems almost impossible to
have an infrastructure network in such environments.
Emergency Operations
MANETs are very useful to be deployed in places that
the conventional infrastructure-based communication
facilities were destroyed by earthquakes, volcanoes, and
any other supernatural situations. This is true since a
MANET is a flexible, mobile, not expensive and can
saves time for deployment phase.
•
Mobile Conferencing
It is unrealistic to expect that all business is done
inside an office environment, so a communication
between a group of people or researchers can be achieved
using MANETs.
•
II. LITERATURE REVIEW
The Binary Exponential Backoff (BEB) [7, 8, 12, 13,
14] is used widely by IEEE 802.11 MAC protocols due to
simplicity and good performance in general. The BEB
algorithm works as the following:
When a node attempts to send a packet to a specified
destination, it first senses the shared medium in order to
decide whether to start transmitting or not. If the channel
is founded to be idle, the transmission process starts.
Otherwise the node should wait a random number of time
between the range of [0, CW-1], this time is calculated
using the formula:
Backoff time = (Rand () MOD CW)*aSlotTime
(1)
After getting the backoff time, the node should wait
until this time reaches zero before start transmitting. The
backoff time (BO) is decremented by one at each idle
© 2012 ACADEMY PUBLISHER
time slot. But if the channel is busy the BO timer will be
frozen. Finally if the node received an acknowledgment
for the packet sent, the contention window (CW) is reset
to minimum for that node. But if the node did not receive
an acknowledgment (send failure occur), the CW is
incremented exponentially to the new backoff value.
S. Manaseer and M. Masadeh [1] proposed the
Pessimistic Linear Exponential Backoff (PLEB). This
algorithm is composed of two increment behaviors for the
backoff value; the exponential and linear increments.
When a transmission failure occurs, the algorithm starts
working by increasing the contention window size
exponentially. And after incrementing the backoff value
for a number of times, it starts increasing the contention
window size linearly. PLEB works the best when
implemented in large network sizes.
S. Manaseer, M. Ould-Khaoua and L. Mackenzie [2]
proposed Fibonacci Increment Backoff (FIB). This
algorithm uses the Fibonacci series formula which is
defined by:
(2)
FIB algorithm aims to reduce the difference between
contention windows sizes generated, resulting in a higher
network throughput than the standard IEEE 802.11.
H. Ki, Choi, S. Choi, M. Chung and T. Lee [15]
proposed the binary negative-exponential backoff
(BNEB) algorithm. This algorithm uses exponential
increments to contention window size during collisions
(transmission failures), and reduces the contention
window size by half after a successful transmission of a
frame. The analytical model and simulation results in [15,
16] showed that the BNEB outperforms the BEB
implemented in standard IEEE 802.11 MAC protocol.
S. Kang, J. Cha and J. Kim [17] proposed the
Estimation-based Backoff Algorithm (EBA). This new
algorithm has two main functions; the first one used to
estimate the number of active nodes, and the second used
to decide which contention window CW is optimal for the
current case. The estimation function uses the average
number of idle slots during backoff time to obtain the
number of nodes which will be after the optimal CW for
the current case.
EBA algorithm outperforms the binary exponential
backoff (BEB), the exponential increase exponential
decrease (EIED), the exponential increase linear decrease
(EILD), the pause count backoff (PCB) and the history
based adaptive backoff (HBAB) in network throughput
and the mean packet delay.
S. Pudasaini, A. Thapa, M. Kang, and S. Shin [18]
proposed an intelligent contention window control
scheme for backoff based on Collision Resolution
Algorithm (CRA). This algorithm keep a history for a
success and failure access attempts in order to use this
history to modify the contention window interval
(CWmin, CWmax). This modification will cause a
dynamic shifting for backoff interval to more suitable
region. This new algorithm made some improvements to
channel efficiency in terms of packet end-to-end delay.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
A. Balador, A. Movaghar, and S. Jabbehdari [19]
proposed a new History Based Contention Window
Control (HBCWC) algorithm for IEEE 802.11 MAC
protocol. HBCWC made an optimization to the
contention window values via saving the last three states
of transmission. The main factor in this algorithm is the
packet lost rate, if this factor increases due to collisions or
channel errors then the CW size will increase. Otherwise
it will decrease.
S. Manaseer and M. Ould-Khaoua [7], proposed the
logarithmic backoff algorithm (LOG) for MAC protocol
in MANETs. This algorithm uses logarithmic increments
to provide a new backoff values instead of exponential
ones. The new backoff values are extracted using the
formula:
(CW) new = (log (CW) old) * (CW) old * aSlotTime.
LOG algorithm generates values that are close to each
other in order to achieve a higher throughput when used
in MANETs.
V. Bharghavan, A. Demers, S. Shenker, and L. Zhang
[10] proposed Multiplicative Increase and Linear
Decrease (MILD) backoff algorithm. This algorithm uses
multiplication by a factor when failed transmission occurs
(due to collision or transmission failure). After a success
transmission occur the contention window CW is
decremented by a factor in order to reduce the probability
of successful users to access the channel all the time. This
decrement helps solving the unfairness problem which
might occur to other users who have collisions and send
failures [8, 9, 10].
J. Deng, P. Varshney, and Z. Haas [9] proposed the
linear multiplicative increase and linear decrease
(LMILD) backoff algorithm. LMILD uses both linear and
multiplicative increments in the case of send failure; that
is when a collision occurs, the colliding nodes increase
their contention window CW multiplicatively, and other
nodes overhearing this collision make a linear increment
to their CW. In the case of successful transmission, all
nodes decrease their contention windows linearly [8, 9].
LMILD has shown a better performance than the
standard IEEE 802.11 when used in large network sizes.
It also outperforms the pessimistic linear
exponential backoff (PLEB) in small networks, but PLEB
achieves better performance than LMILD in large
network sizes [1].
III. THE NEW PROPOSED BACKOFF ALGORITHM
A. Overview
In general, backoff algorithms tend to increase the
contention window (CW) size after each transmission
failure. Since this is true, a backoff algorithm should use
a suitable increment for CW size in order to achieve the
best performance. Many increment behaviors were used
in this field such as: linear, exponential, logarithmic, and
Fibonacci series. If we split the networks into three types:
small, medium and large, each increment scheme would
suit at the most two network types and drops dramatically
at the third one. For example, the exponential increment
of BEB algorithm which is used in standard IEEE 802.11
© 2012 ACADEMY PUBLISHER
157
MAC does not achieve the best performance due to large
CW gaps produced. Another example is a linear
increment of LMILD; it does not allow adequate backoff
time before data retransmission.
The Binary Exponential Backoff (BEB) algorithm
increases the CW sizes exponentially based on
transmission failure. That is, when a node has a packet to
transmit, first it starts sensing the channel. If the
channel is found to be idle, the node starts transmitting
immediately the data packets. Otherwise, the backoff
mechanism is triggered. Furthermore, a backoff timer is
selected randomly from the current CW size; this timer is
decremented only at each time slot found to be idle.
When the timer reaches zero, the node transmits the data
packets. If the acknowledgement received from the
destination node, then CW size is reset to minimum. On
the other hand, if the acknowledgement is lost the CW
size is incremented exponentially. See Fig. 3.
Figure 3: BEB algorithm description
Another backoff algorithm called pessimistic linear
exponential backoff (PLEB). This algorithm combines
two increment behaviors: exponential and linear
increments. PLEB assumes that congestion in the
network will not be resolved in the near future. Thus,
PLEB selects exponential increments at the first stages of
the algorithm and continues to linear behavior. See Fig. 4.
The next section suggests a new backoff algorithm that
aims to improve the network performance overall. Our
goal is to achieve a higher data delivery ratio with less
overhead to the network. The new suggested algorithm
uses a combination of different increments in order to
take the advantage of each. The increments used in this
algorithm are: linear, logarithmic, and exponential. The
exponential increment aims to produce adequate CW
lengths at the first stages of the new algorithm.
Logarithmic increment provides proper increments to
CW if it is still small; generally logarithmic increment is
used as a transition stage towards linear increment to
avoid continues exponential increments. The last
increment used is linear increment, which aims to avoid
158
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
large increments of exponential and logarithmic ones.
The simulation results presented in the next paper show
that the new proposed backoff algorithm improves the
network delivery ratio and overhead.
After that, the lines 8 and 9 shows that SABA calculates
the average of CW sizes in the history array in order to
start a new increment behavior based on the average
value. This average is computed only once. Now, the
lines 11-16 express that one increment behavior will be
chosen: linear or logarithmic. If the average is not high
(less than threshold N) the next increment behavior will
be the logarithmic increment. Otherwise it will be the
linear increment.
1 Set BO to initial value
2 While Bo ≠0 do
3 For each time slot
4
If channel is idle then BO=BO -1
5
6
7
Figure 4: PLEB algorithm description
B. The Smart Adaptive Backoff Algorithm
The Smart Adaptive Backoff Algorithm (SABA) is the
new suggested backoff algorithm. This algorithm
assumes that a network performance can be enhanced if
the network is very sparse or congested too much. The
explanation for this is the following; SABA assumes that
solving the network collision in very sparse and
congested networks will not be in the near future.
Therefore, exponential increments are used. In very
sparse networks, paths can be broken easily due to
mobility, and mostly there exists only one path in the
route table. So, the exponential increments will provide
adequate values to establish other paths and start using
them. On the other hand, in congested networks many
nodes in the network use the same paths, so we should
provide more time before start transmitting on this path.
However, figure 5 shows in detail the basic functionality
of SABA algorithm. As shown, the first 5 lines of the
algorithm set the initial value for the backoff timer. And
then, starts the decrement of that time based on idle time
slots. This means, the timer will be decremented only at
idle time slots. Otherwise, it freezes. After the timer
reaches zero, the data packet is transmitted. Now, in case
of successful transmission the CW value is saved in the
history array as shown in line 21 of the algorithm.
Otherwise, the backoff mechanism will be triggered. The
lines 7-19 show a brief description of the adaptive
process in SABA algorithm. It starts with line 7 by
incrementing the CW exponentially only for a number of
times based on transmission success. In other words,
SABA provides an exponential increment for a node, and
saves the CW size in case of transmission success. This
process is repeated until the array of five elements is full.
© 2012 ACADEMY PUBLISHER
Wait for a period of DIFS then Send
If (Send-Failure) then
If (CW array of last five successes is full)
8
If (Array used for the first time)
9
Calculate the average of the history array and use
it as a new CW value
Else
10
11
If (CW > N) then
12
CW = CW + T
13
Backoff-Timer = Random x;
14
Else
15
CW = Log (CW) * CW
16
Backoff-Timer = Random x;
17
1 ≤ x ≤ CW - 1
1 ≤ x ≤ CW - 1
Else
18
CW = CW * 2
19
1 Set BO to initial value
20 Else
21
Save the CW value used in the history
22 Go to line number 1
23 Stop Backoff-Timer = Random x
1 ≤ x ≤ CW - 1
20 Else
21
Save the CW value used in the history
22 Go to line number 1
23 Stop
Figure 5: The Smart Adaptive Backoff Algorithm
The purpose of using the average is to reduce the CW
size if it is a large number. Both of logarithmic and linear
increments aim to avoid excessive of CW sizes in order
to enhance the network performance.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
A. Simulation Environment
Our simulations were run using a network of 10, 20, 50
and 100 nodes placed randomly within a 1000 meter ×
1000 meter area. Each node has a radio propagation
range of 250 meters, and the channel capacity is 2 Mb/s.
Each run executed for 900 seconds. We used the IEEE
802.11 as the MAC layer protocol. The Constant Bit Rate
(CBR) node traffic is used in the simulations. We used
the random waypoint model for node mobility. We used
various node maximum speeds: 1, 2, 3 and 4 meter per
second. In addition, we used traffic loads of 5, 10 and 20
packets per second, repeated for 5 and 10 sources.
B. Simulation Results
Different performance metrics were used to evaluate
backoff algorithms [1, 2, 4, 6, 7, 9, 10, 11, 12]. This study
uses data delivery ratio and network overhead parameters
to measure performance levels. The ideal case is to
achieve maximum delivery ratio and minimum network
overhead.
100
Delivery Ratio (%)
In this research paper the network performance is
measured by two criteria: packet delivery ratio and
network overhead. In this paper, we present and evaluate
the simulation results that were obtained for different
scenarios. In the simulation experiments, we varied the
number of sources and the maximum node speed.
We implemented the proposed backoff algorithm
SABA using the GloMoSim [21] simulator to evaluate
the performance of the new algorithm compared to the
well-known BEB and PLEB algorithms.
transmissions. After that, if the result is a large contention
window size, the increment is linear. Otherwise,
increment to contention window is logarithmic and then
continues in linear way. The logarithmic increment is
significantly smaller than exponential increment but it
still larger than linear (this is needed in large networks to
perform in a better way).
80
60
BEB
40
PLEB
20
SABA
0
1
2
© 2012 ACADEMY PUBLISHER
4
Mobility Speed(m/sec)
0.5
0.4
0.3
BEB
0.2
PLEB
0.1
SABA
0
1
2
3
4
Mobility Speed (m/sec)
Figure 7: Overhead of BEB, PLEB and SABA for 20 nodes, 5
sources each source sending 20 packets per second.
100
Delivery Ratio (%)
The figures (6, 7, 8, 9, 10, 11, 12, 13) show that SABA
achieves the best performance in light and heavy (small
and large) networks. This is not surprising since a
network with a small number of nodes does not
frequently trigger backoff mechanism; therefore, SABA
which uses the average of last five exponential
increments (successful increments only) outperforms the
exponential increments in BEB and PLEB. Moreover,
when the number of nodes increases and the network
become of a large size (e.g. a network of 100 nodes), the
exponential increments has essential role in network
performance; that is, in BEB the exponential increments
continues to enlarge the gaps between contention window
sizes which in turn significantly reduces the network
performance. Moreover, in PLEB algorithm the linear
increments begin early in a way that does not allow the
increments to the best values for large networks (i.e. the
increments still small compared to network size and
nodes mobility speed). On the other hand, SABA
provides the best contention window increments. SABA
aims to utilize exponential, logarithmic and linear
increments to achieve the best performance in large
networks. It starts with exponential increments until the
node successfully sends five times. After that, it is
expected to have a large contention window size.
Therefore, it is reduced by calculating the average
contention window sizes of these successive
3
Figure 6: Data delivery ratio of BEB, PLEB and SABA for 20 nodes,
5 sources each source sending 20 packets per second.
Overhead
IV. SIMULATION AND RESULTS' EVALUATION
159
80
60
BEB
40
PLEB
20
SABA
0
1
2
3
4
Mobility Speed (m/sec.)
Figure 8: Data delivery ratio of BEB, PLEB and SABA for 20 nodes,
10 sources each source sending 10 packets per second.
At medium sized networks, the simulation results have
shown in figures (14, 15, 16, 17) that SABA is still
closely comparable the BEB and PLEB algorithms, and
even gives a better performance at high traffic rates.
BEB, PLEB and SABA backoff algorithms start to
increase the contention window size exponentially.
160
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
100
0.3
0.2
Delivery Ratio(%)
Overhead
0.25
BEB
0.15
PLEB
0.1
0.05
SABA
80
60
BEB
40
PLEB
20
0
SABA
0
1
2
3
4
1
2
3
4
Mobility Speed (m/sec.)
Mobility Speed (m/sec.)
Figure 9: Overhead of BEB, PLEB and SABA for 20 nodes, 10
sources each source sending 10 packets per second.
Figure 12: delivery ratio of BEB, PLEB and SABA for 100 nodes, 10
sources each source sending 20 packets per second.
50
40
80
60
Overhead
Delivery Ratio(%)
100
BEB
40
20
30
BEB
20
PLEB
PLEB
10
SABA
0
0
1
1
2
3
4
2
3
4
Mobility Speed (m/sec.)
Mobility Speed (m/sec.)
Figure 10: Data delivery ratio of BEB, PLEB and SABA for 100 nodes,
5 sources each source sending 20 packets per second.
Figure 13: Overhead of BEB, PLEB and SABA for 100 nodes, 10
sources each source sending 20 packets per second.
.
2.5
2
BEB
1.5
PLEB
1
Delivery Ratio (%)
100
3
Overhead
SABA
80
60
BEB
40
PLEB
20
SABA
0
0.5
SABA
1
0
1
2
3
4
Mobility Speed (m/sec.)
© 2012 ACADEMY PUBLISHER
3
4
Mobility Speed (m/sec.)
Figure 14: Data delivery ratio of BEB, PLEB and SABA for 50 nodes, 5
sources each source sending 20 packets per second.
Overhead
Figure 11: Overhead of BEB, PLEB and SABA for 100 nodes, 5
sources each source sending 20 packets per second
The number of backoff processes expected to be
moderate. Therefore, the continuous exponential
increments in BEB algorithm would not be a problem in
this case. Moreover, SABA algorithm continues with
logarithmic increments which are more suitable for high
traffic rates in this type of networks. While the PLEB
algorithm starts linear increments early making this
mechanism unsuitable to gain the best performance for
medium networks.
2
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
BEB
PLEB
SABA
1
2
3
4
Mobility Speed (m/sec.)
Figure 15: Overhead of BEB, PLEB and SABA for 50 nodes, 5
sources each source sending 20 packets per second.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Delivery Ratio (%)
100
V. CONCLUSIONS
80
60
BEB
40
PLEB
20
SABA
0
1
2
3
4
Mobility Speed (m/sec.)
Figure 16: Data delivery ratio of BEB, PLEB and SABA for 50 nodes,
10 sources each source sending 10 packets per second.
Overhead
161
3.5
3
2.5
2
1.5
1
0.5
0
BEB
PLEB
SABA
1
2
3
4
Mobility Speed (m/sec.)
Figure 17: Overhead of BEB, PLEB and SABA for 50 nodes, 10
sources each source sending 10 packets per second
Furthermore, the results show some other important
information; first, the linear increment behavior directly
affects the network overhead. For example, PLEB
algorithm provides the lowest network performance
compared to BEB and SABA for all scenarios. This is
true since the linear increments in the early stages of
PLEB do not allow adequate increments for the CW
values. To explain this in detail, the three types of
network should be studied extensively. That is, in a
sparse network linear increments make the source node
sends the route request more frequent in order to establish
a path between source and destination nodes (this is hard
because the number of nodes is still small). Furthermore,
in a medium and large network it is expected to have a
more frequent backoff triggers. Therefore, it is normal to
have a congested network and more broken routes. For all
reasons mentioned above, it is normal to expect that
linear increments in backoff algorithms cause a higher
network overhead than exponential ones for the applied
scenarios in this thesis. Secondly, in case of lightly
loaded networks, the data delivery ratio is increasing.
This means that a network of 20 nodes (5 and 10 sources
sending data packets) and a network of 50 nodes (only 5
sources sending data packets) the data delivery ratio has
increased due to the increasing mobility of nodes can be
more useful. In other words, the data delivery ratio
increases because when routes break due to mobility
some other routes are built quickly. Finally, at highly
loaded networks (ex. network of 100 nodes) the data
delivery ratio decreases as mobility speed increases. This
© 2012 ACADEMY PUBLISHER
In this paper we presented a new backoff algorithm for
MANETs called the Smart Adaptive Backoff Algorithm
(SABA). The main objective of this work is to evaluate
the performance of the new backoff algorithm in terms of
network size, mobility speeds and traffic rates. The
results obtained approve that changes made to contention
window size increment and decrement directly affects
network performance metrics such as data delivery ratio
and overhead.
The results have shown that SABA outperforms BEB
and PLEB algorithms in different network types. The data
packet delivery ratio of SABA against BEB and PLEB
algorithms was fluctuating. For example, in a small
network type when a transmission rate is twenty packets
per second, the number of sources is five and mobility is
low SABA outperform BEB and PLEB by 37.11% and
38.45%, respectively. At high mobility, SABA
outperforms BEB and PLEB by 7.74% and 14.13%,
respectively. At medium network type, when a
transmission rate is ten packets per second, the number of
sources is ten and mobility is low SABA outperforms
BEB and PLEB by 5.59% and 22.79%, respectively. At
high mobility, SABA outperforms BEB and PLEB by
13.19% and 21.56%, respectively. At large networks,
when a transmission rate is ten packets per second, the
number of sources is ten and mobility is low SABA
outperforms BEB and PLEB by 0.93% and 30.34%,
respectively. At high mobility, SABA outperforms BEB
and PLEB by 36.42% and 39.67%, respectively.
The network overhead of SABA against BEB and
PLEB algorithms was fluctuating. For example, in a
small network type when a transmission rate is twenty
packets per second, the number of sources is five and
mobility is low SABA outperform BEB and PLEB by
0.249% and 0.392%, respectively. At high mobility,
SABA outperforms BEB and PLEB by 0.299% and
0.746%, respectively. At medium network type, when a
transmission rate is twenty packets per second, the
number of sources is ten and mobility is low SABA
outperforms BEB and PLEB by 0.427% and 8.875%,
respectively. At high mobility, SABA outperforms BEB
and PLEB by 1.793% and 28.486%, respectively. At
large networks, when a transmission rate is twenty
packets per second, the number of sources is ten and
mobility is low SABA outperforms BEB and PLEB by
0.454% and 6.645%, respectively. At high mobility,
SABA outperforms BEB and PLEB by 18.248% and
39.211%, respectively.
In general, the results of this research paper indicate
that each type of networks needs a different way to
handle contention window increment. That is, for small
networks low increments are preferred (lower than
exponential). On the other hand, for medium and large
networks it is preferred to have low increments after large
ones.
Finally, this work has studied the effect of choosing
the behavior changing point between linear, logarithmic
162
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
and exponential increments in the proposed algorithm
SABA. Results have shown that using the suitable
increment type according to the network status increases
overall network performance.
ACKNOWLEDGMENT
We would like to thank the staff of the Faculty of
Computer and Information Technology, Jordan
University of Science and Technology (JUST) and the
University of Jordan (JU), Jordan, for their great and
continuous support.
REFERENCES
[1] S. Manaseer and M. Masadeh. “Pessimistic Backoff for
Mobile Ad hoc Networks”. Al-Zaytoonah University, the
International
Conference
on
Information
Technology (ICIT'09), Jordan, 2009.
[2] S. Manaseer, M. Ould-Khaoua, and L. Mackenzie.
“Fibonacci Backoff Algorithm for Mobile Ad Hoc
Networks”. Liverpool John Moores University, the 7th
Annual Postgraduate Symposium on the Convergence of
Telecommunications, Networking and Broadcasting
(PGNET 06), Liverpool, 2006.
[3] Y. Yuan, A. Agrawala. “A Secure Service Discovery
Protocol for MANET”. Computer Science Technical
Report CS-TR-4498, Computer Science Department,
University of Maryland, 2003.
[4] S. Manaseer, M. Ould-Khaoua, and L. Mackenzie. “On a
Modified Backoff Algorithm for MAC Protocol in
MANETs”. Int. J. of Information Technology and Web
Engineering, Vol. 2(1), pp. 34-46, 2007.
[5] P. Karn. “MACA - A new channel access method for
packet radio”. ARRL/CRRL Amateur Radio 9th computer
Networking Conference, London, pp. 134–140, 1990.
[6] M. Bani Yassein, S. Manaseer, and A. Al-Turani. “A
Performance Comparison of Different Backoff Algorithms
under Different Rebroadcast Probabilities for MANET’s”.
University of Leeds, the 25th UK Performance
Engineering Workshop (UKPEW), UK, 2009.
[7] S. Manaseer and M. Ould-Khaoua. “Logarithmic Based
Backoff Algorithm for MAC Protocol in MANETs”.
Technical Report, University of Glasgow, 2006.
[8] H. Wu and Y. Pan. “Medium Access Control in Wireless
Networks”. Nova Science Publishers Inc, pp. 26- 29,
2008.
[9] J. Deng, P. Varshney, and Z. Haas. “A New Backoff
Algorithm for the IEEE 802.11 Distributed Coordination
Function”. Communication Networks and Distributed
Systems Modeling and Simulation (CNDS’04), San Diego,
California, 2004.
[10] V. Bharghavan et al, “MACAW: A Media Access Protocol
for Wireless LAN’s”, in Proc. ACM SIGCOMM ’94, pp.
212–225, 1994.
[11] S. Zhalehpoor and H. Shahhoseini. “SBA Backoff
Algorithm to Enhance the Quality of Service in
MANETs”. International Conference on Signal Acquisition
and Processing ICSAP, Kuala Lumpur, Malaysia, pp. 4347, 2009.
[12] C. Huy, H. Kimz, and J. Hou. “An Evaluation of the
Binary Exponential Backoff Algorithm in Distributed
MAC Protocols”. Technical Report, University of Illinois –
Urbana Champaign, 2005.
© 2012 ACADEMY PUBLISHER
[13] J. Goodman et al. “Stability of binary exponential
backoff”. In the Proc. of the 17-th Annual in Journal of the
ACM, Vol. 35(3), pp. 579–602, 1988.
[14] Y. Wang, “Medium Access Control in Ad Hoc Networks
with Omni-directional and Directional Antennas”,
University of California, doctoral Dissertation, 2003.
[15] H. Ki, Choi, S. Choi, M. Chung, and T. Lee. “Performance
evaluation of Binary Negative-Exponential Backoff
Algorithm in IEEE 802.11 WLAN”. Lecture Notes in
Computer Science, Springer-Verlag Berlin Heidelberg, vol.
4325, pp. 294–303, 2006.
[16] B. Choi, S. Bae, T. Lee, and M. Chung. “Performance
Evaluation of Binary Negative-Exponential Backoff
Algorithm in IEEE 802.11a WLAN under Erroneous
Channel Condition”. Lecture Notes in Computer Science,
Springer-Verlag Berlin Heidelberg, vol. 5593, pp. 237-249,
2009.
[17] S. Kang, J. Cha, and J. Kim. “A Novel Estimation-Based
Backoff Algorithm in the IEEE 802.11 Based Wireless
Network”. 7th Annual IEEE Consumer Communications &
Networking Conference (CCNC), Las Vegas, Nevada,
USA, 2010.
[18] S. Pudasaini, A. Thapa, M. Kang, and S. Shin. “An
Intelligent Contention Window Control Scheme for
Distributed Medium Access”. 7th Annual IEEE Consumer
Communications & Networking Conference (CCNC), Las
Vegas, Nevada, USA, 2010.
[19] A. Balador, A. Movaghar, and S. Jabbehdari. “History
Based Contention Window Control in IEEE 802.11 MAC
Protocol in Error Prone Channel”. Journal of Computer
Science, 6 (2), pp. 205-209, 2010.
[20] B. Ramesh, D. Manjula. “An Adaptive Congestion Control
Mechanism for Streaming Multimedia in Mobile Ad-hoc
Networks”. International Journal of Computer Science and
Network Security (IJCSNS), Vol. 7(6), pp. 290-295, 2007.
[21] X. Zeng, R. Bagrodia, M. Gerla. “GloMoSim: A Library
for Parallel Simulation of Large-scale Wireless Networks”.
Proceedings of the 12th workshop on Parallel and
distributed simulation, Banff, Alberta, Canada, pp. 154161, 1998.
[22] I. Stojmenovic. “Handbook of wireless networks and
mobile computing”, Wiley, New York, 2002.
[23] C-K Toh. “Ad hoc mobile wireless networks, protocols and
systems”, Prentice-Hall, New York, 2002.
[24] IEEE, ANSI/IEEE standard 802.11, 1999 Edition (R2003),
Part 11: “Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) Specifications”.
[25] R. Rajaraman. “Topology control and routing in ad hoc
networks: a survey”. ACM Special Interest Group on
Algorithms and Computation Theory (SIGACT), Vol.
33(2), 2002.
[26] V. Rodoplu and T. Meng. “Minimum energy mobile
wireless networks”. IEEE Journal of Selected Areas in
Communications, Vol. 17(8), pp. 1333-1344, 1999.
[27] S. Nesargi and R. Prakash. “MANETconf: configuration of
hosts in a mobile ad hoc network”. Proceedings of IEEE
Twenty-First Annual Joint Conference of the IEEE
Computer and Communications Societies (INFOCOM),
Vol(2), pp. 1059-1068, 2002.
[28] Y. Tseng and T. Hsieh. “Fully power-aware and locationaware protocols for wireless multi-hop ad hoc networks”.
Proceedings of the Eleventh International Conference on
Computer Communications and Networks, pp. 608-613,
2002.
29. Manaseer. “On Backoff Mechanisms for Wireless Mobile
Ad Hoc Networks”. Doctoral Dissertation, United
Kingdom (UK), Glasgow University, December, 2009
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Muneer O. Bani Yassein received his
B.Sc. degree in Computing Science and
Mathematics from Yarmouk University,
Jordan in 1985 and M. Sc. in Computer
Science, from Al Al-bayt University,
Jordan in 2001. He received his Ph.D.
degree in Computing Science from the
University of Glasgow, U.K. in 2007.
His current research interests include
networking protocol development and analysis the performance
probabilistic flooding in mobile ad hoc networks.
Saher S. Manaseer received PhD degree
in Computer Science from the
Department of Computing Science at the
University of Glasgow in 2009. His main
area of research is Computer Networks
and Embedded Systems. Currently, he is
an active researcher in the field of
Mobile Ad Hoc Networks. More
specifically, his research on MANETs is
focused on developing MAC layer protocols. Before obtaining
his PhD, He got his Masters in Computing Science from the
University of Jordan in 2004 where his research in that era was
focused on Software Engineering. Since he earned his
undergraduate degree in Computer science, he has worked in
two parallel directions, being Training and software
development.
Ahmad A. Momani received his B.Sc.
degree in Computer Science from Jordan
University of Science and Technology,
Jordan in 2007 and M. Sc. in Computer
Science, from Jordan University of
Science and Technology, Jordan in 2011.
His main area of research is Computer
Networks. Currently, he is an active
researcher in the field of Mobile Ad Hoc
Networks. More specifically, his research on MANETs is
focused on developing MAC layer protocols. Before obtaining
his Masters, he has worked as a programmer and then a teacher
in the Ministry of Education.
.
© 2012 ACADEMY PUBLISHER
163
164
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
ConTest: A GUI-Based Tool to Manage InternetScale Experiments over PlanetLab
Basheer Al-Duwairi
Jordan University of Science and Technology/Department of Network Engineering and Security, Irbid, Jordan
Email: basheer@just.edu.jo
Mohammad Marei, Malek Ireksossi, and Belal Rehani
Jordan University of Science and Technology/Department of Computer Engineering, Irbid, Jordan
Email: {mymarei, malek.erq, belalrehani}@hotmail.com
Abstract— PlanetLab is being used extensively to conduct
experiments, implement and study large number of
applications and protocols in Internet-like environment.
With this increase in PlanetLab usage, there is a pressing
need to have efficient way to setup and mange experiments.
This paper proposes a graphical user interface-based tool,
called ConTest, to setup and mange experiments over
PlanetLab. ConTest enables PlanetLab users to setup
experiment and collect results in a transparent and easy
way. This tool also allows different measurements for
different variables over the PlanetLab network. The paper
discusses the design and implementation of ConTes and
shows different scenarios.
Index Terms—Internet measurements, PlanetLab. API,
Networking, GUI.
I. INTRODUCTION
PlanetLab is an open shared platform for developing,
deploying, and accessing planetary scale applications [1].
Currently, it has more than 1100 nodes distributed all
over the world. It allows its users to freely access the
shared nodes, upload programs and execute them. It is
used by many distributed applications developers all
around the world for testing their applications (e.g., [2, 3,
4, 5]). It can also serve the purpose of having machines
distributed all around the Internet to break the limit of
locality and allow the users to perform their
measurements on random nodes at random places to
ensure generality and confidence in the obtained results.
PlanetLab is being used extensively to conduct
experiments, implement and study large number of
applications and protocols in Internet-like environment.
These applications and protocols fall into different areas
that include real-time measurements of delay, bandwidth,
and security. With this increase in PlanetLab usage, there
is a pressing need to have efficient way to setup and
mange experiments. Having such tool would be a great
assist not only for Planetlab community, but also to new
researchers/users who find it difficult to explore and
utilize this platform. In fact, several systems (as described
in Section V) have been proposed to achieve that goal.
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.164-171
This paper builds on previous efforts in this direction and
proposes a GUI-based tool, called ConTest, that allows
users to visually manage experiments in Internet-like
environment. In this regard, ConTest provides great deal
of flexibility and simplicity to conduct experiments over
PlanetLab. Using this tool, users have the ability to
visually create the network topology they want simply by
selecting PlanetLab nodes as they appear on the world
map view. Also, it allows users to specify the role of
each node (client, server, or proxy) and to generate
different type of traffic (e.g., TCP, UDP, ICMP, etc.). In
addition, ConTest has the capability to perform basic
operations in sequence or in parallel. This include
authentication, file upload, and command execution.
ConTest provides efficient and simple way to deploy and
test applications over Planetlab. With this tool it becomes
very easy to study the properties of Internet applications
and protocols. It also becomes easy to measure their
performance in terms of response time, delay jitter,
packet loss, etc. With its ability in supporting application
deployment and monitoring in Internet-Like environment
such as PlanetLab, we believe that ConTest would
promote research in this field and would significantly
contribute to the increase of PlanetLab users because of
its attractive features that hide many of the complications
behind the scene.
The rest of this paper is organized as follows: Section II
explains ConTest architecture. Section III discusses
implementation details. Section VI discusses security
issues. Section V discusses related work. Finally,
conclusion and future work is presented in Section VI.
II. CONTEST ARCHITICURE
The design and implementation of ConTest is based on
modularity and usability concepts. In this regard, ConTest
is composed of several modules with each module having
specific functionality. In particular, ConTest has the
following modules: Authentication Module, Command
execution module, File management module, and
Connection management module. What follows is a
description of each module.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
A. Authentication Module
Typically, users access PlanetLab nodes through slices.
Basically, a slice is a multi-user account registered on the
PlanetLab Central (PLC) to offer all the users of that
account the same resources on the nodes which they add
to their slice such that users of that account work on the
same project and preserve node’ resources of that slice.
Accessing a PlanetLab node requires the user to have
access to the slice to which the node belongs to.
PlanetLab offers a web interface for the users to allow
them to access their slices and see the nodes registered
there. But it also offers an API that allows external
programs to communicate with PLC databases through
remote procedure calls (RPC) to retrieve users' account
and node information. ConTest utilizes this API to
perform authentication.
The authentication module is a program that
communicates directly with the PlanetLab central API. It
accepts the authentication information (email, PlanetLab
password and slice name) from the main module. This
information is then validated by the PLC. After that, it
retrieves the slice info from the PLC and the node list
registered on this slice, with all the necessary data about
each node (ID, name, location, status, etc.) as illustrated
in Fig. 1.
165
connects to the desired nodes using the secure public key
cryptography authorization, sends the command to the
destination nodes and retrieves the output and error
messages from each node.
C. File Management Module
Similar to the command execution module, file copying is
done through SSH and using the secure copy function.
After accepting the key pair identification, the files are
uploaded to the destination nodes and their console dump
is retrieved to ensure that the copying process was
successful.
D. Connection Management Module
This module represents the operational core of the tool. In
order for the user to perform their tests, they need to
connect the nodes together using a testing program which
they also need to upload. The connection management
module does that for the user. Also, this module heavily
utilizes the program capabilities. It uses a special class to
setup the connections required by the user. The role of
each node (e.g., client, server) is determined using the
programmed connection algorithm. The connection
management module communicates with the map
interface on the user interface to help the user select their
connection topology in a simple and efficient way. It also
utilizes the file copying and command execution modules
to upload the test files to the desired nodes, execute these
files, retrieve the results from each node and output these
results to the GUI.
The results associated with an experiment are stored to
separate files for reference. Also some statistics (e.g.,
bandwidth, round trip time, etc.) are displayed to the
users upon the execution of these experiments. Fig 2.
shows the execution flow for the connection management
module and its interaction with other modules in each
step.
Figure 1: How the authentication module accesses the
API to retrieve node list
B. Command Execution Module
The command execution module allows users to run their
own custom commands on the nodes to which they have
access (i.e., within their slices). Also, it is used in the
connection management module to run the tests requested
by the user. In order for the tool to communicate with the
remote nodes and access their resources to perform the
required test, and due to the security restrictions on the
access to PlanetLab nodes, any connection to the nodes
must be established through the secure shell (SSH)
protocol. Therefore, command execution is done through
SSH connections, where the command execution module
© 2012 ACADEMY PUBLISHER
Figure 2: Execution flow in the connection
management module
E. The Graphical User Interface (GUI) module.
166
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
The GUI represents the main program that links all other
modules of the tool. The GUI module accepts user inputs
and manages outputting results and error messages to the
user in text file format and to the program log. It
communicates with the authentication module to read
node data and display them in table format, as well as
identify and show their location on the world map using
the retrieved coordinates from the authentication module.
It also sends the user input to the command execution and
file management modules and retrieves the console log
from each module to display it in the program log or save
it to separate files. Also, it integrates the connection
management module with the command execution/file
management modules to achieve the tests required by the
user.
Node view class: This class is mainly used to represent
the nodes graphically, it also stores all the available data
of the nodes, namely:
• Node ID.
• Node name.
• Location.
• Status.
• Coordinates (for display in the GUI's map view).
It also contains pointers to other nodes. These pointers
are used in creating linked lists that help in constructing
the required connections for the tests performed by the
connection management module. Fig. 3 illustrates the
relationship between the different classes; it also shows
the relationship between the main module and the
authentication module.
F. ConTest Classes
GUI class: The core of ConTest is the GUI module/class,
which connects the other modules together and manages
their interactions. This class is responsible for interacting
with the user, I/O management between the sub-classes,
and error handling. Most of the content of the main class
concerns the graphical widgets displayed on the GUI, the
initiating of the other modules and the management of
their inputs and outputs. So we won't go deep into
describing this class as it contains many auto generated
elements (the graphical widgets). That leaves us with the
sub-classes in the design, which are:
Copying thread class: The copying thread class is the
class that represents the file management module.
Instances of this class receive and store the essential data
of the nodes to which the copying is going to be
performed, as well as the data of the file to be copied to
the nodes, this includes:
• Node name.
• Key file (for the ssh connection).
• Source file (the file to be copied).
• Destination directory on the remote node.
The instances run a scp (secure copy) command using the
information passed by the main module, retrieves the
console output from the remote nodes and relay this
output to the GUI.
Execution thread class: Similar to the file copying class,
instances of this class contain the data of the nodes which
the command is going to be executed on:
• Node name.
• Key file.
• Command string.
Each instance of this class runs a ssh command using that
information, which is passed to it by the main module,
retrieves the console output from the remote nodes and
relays this output back to the GUI.
© 2012 ACADEMY PUBLISHER
Figure 3: Relationship between different classes
III. IMPLEMENTATION DETAILS
A. Tools and Programs used to Implement ConTest
ConTest was implemented using Qt Qreator [6][7], a
widely used IDE (integrated development environment)
that helps programmers create powerful graphical tools,
which provides an effective way to integrate an easy to
use, simple graphical interface with the ability to perform
a wide set of different operations (like accessing APIs,
using the SSH commands, displaying map, etc.). Qt
Qreator has a huge library of graphical and non graphical
classes and functions that enable the programmer to do
almost anything they want, helping them give their
programs the maximum functionality. It uses C++ as the
programming language, and its set of libraries makes the
programming process fast and effective, and helps the
programmer to focus on the core functionality of their
program while it handles the graphical aspects.
In addition Ping [8, 9] and Iperf [10] have been integrated
within ConTest to provide means to measure different
aspects of network performance related metrics, such as
round trip time and connection throughput. In ConTest,
this can be done by starting two processes on the
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
designated remote node, one measures the bandwidth
using Iperf tool and the other measures the round trip
time using the ping command. These processes then
return their output through SSH to the program in order
for it to record it locally and analyze it to get the required
connectivity statistics about all the nodes of interest to the
user.
167
handler or, as it is called in the Qt environment,
signal/slot pair. When a thread has some output that it
needs to write to the interface, it emits a signal to the GUI
telling it that its output is ready. The GUI receives this
signal through a specific slot and puts it in the queue of
signals that wish to access the interface. This regulation
prevents multiple access to one object at the same time
and gives each thread its turn to write to the interface.
B. Inter-process Communication
The main program, the GUI, constitutes the main process
that initializes and connects other processes, thus the
other sub-processes take their required running data from
the GUI through an inter-process link, then they run their
separate codes independently in the background and
return the output to the main program through the same
link.
When created, each sub-process gets initialized and sent
its parameters, and then each process establishes a
connection to the main process through two channels, the
standard output channel (stdout), which communicates all
outputs from this process to the main process, and the
standard error channel (stderr) which transfers the error
messages (if any) from the sub-process back to the calling
process. Even when a sub-process needs to communicate
with another sub-process, this communication needs to go
through the main process as illustrated in Fig. 4.
Figure 4: Process communication through channels
C. Multi-threading
Since the execution of commands and file copying
through the SSH is a lengthy process, and since we need
to do a lot of copying, command execution and file
writing, along with the active interaction with the GUI,
we had to program the tool in the form of threads.
Threads are, essentially, child processes that execute
concurrently. Each separate thread runs in parallel with
the GUI, thus keeping the program responsive at all times
and increasing the utilization of the machine resources.
Of course, these threads need to access the GUI to write
their output through it to the user, however, multiple
threads writing to the same object (i.e. the program log) is
prohibited as it can cause access errors and even program
crashes. So we had to use the concept of event/event
© 2012 ACADEMY PUBLISHER
D. Authentication
As mentioned earlier, the PlanetLab Central has high
security standards that restricts the access to their user
and node databases, but fortunately, they provide the API
to access their PLC database and retrieve the node data
we require for our tool, thus we have programmed the
separate authentication script. This script does the
authentication through the API functions and using the
user's data provided in the GUI (their slice name and
PlanetLab registered email and password), and retrieves
the detailed node list for that specific user. Also, the
connection to the nodes themselves requires the use of the
SSH protocol, and since SSH uses public key
cryptography to authenticate to the remote computer, we
had to ask the user for the key file which they have
registered in their PlanetLab account.
Another issue with the SSH connections is that SSH
protocol requires the key's password each time you
connect to the remote host for confirmation. This is the
default process and it is recommended as it prevents
altering or theft the user's SSH key, but if the user needs
to access multiple hosts at a time, it will make the
connection process more redundant and time consuming.
However, since this check can be disabled through the
command options, we decided to give the user the liberty
to choose whether to enable this confirmation or not, but
disabling the check is not recommended as it may raise
security warnings and risk the user's key security.
VI. CONTEST FUNCTIONALITY AND USAGE
ConTest provides an easy and flexible way to setup and
perform experiments over PlanetLab. This is done mainly
through the experiment manager interface, where the user
specifies experiment parameters and through the
connection tester interface which establishes different
connection schemes between nodes and measures some
aspects of these connections. Before going into the
details of these two interfaces we describe the way
authentication is performed.
A. Authentication
The authentication dialog pops up at the start-up of the
program, it requires the following data to enable you to
use the device:
• Email: the e-mail which you have registered in your
PlanetLab central account.
168
•
•
•
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Password: the password of your account.
Slice name: the name of a slice reserved for you on
PlanetLab.
Key file: the directory of your private key that is
used to access the nodes through the secure shell
(ssh).
Upon a successful authentication, user information will
be displayed in the upper right corner; the number of
nodes in the provided slice will be displayed in the
program log. Also the details of the node list registered
on the slice are displayed in the table view and the nodes
themselves are represented on the map as shown in Fig.5.
[ Figure 8: Authentication successful ]
Figure 5: After successful authentication
Connection testing tab provides an interface for users that
allows them to test the connection between any nodes
from the list of nodes within a slice. It also allows
selecting the protocol of preference. The test results are
displayed on the program log and stored to external files
for documentation. The basic functions that can be
performed within this tab are:
•
•
Test Protocol: from here you should be able to select
your has its own client, server and forwarding
program. You can select the protocol from the “test
protocol” drop down box. And you can also browse
for your own programs if you have customized a
code to be used with ConTest. To restore the defaults
test programs, just press the “restore defaults”
button.
The Map Interface: In the connection tester, you can
add nodes to test only through the map interface.
This is done by clicking on a node using the left
mouse button. As long as you keep selecting nodes
using the left button you will keep the path open.
Until you close it by clicking a node using the right
mouse button, thus setting it as a server. Currently,
you can’t give a single node two roles (e.g. a client
and a server at the same time).
After setting up your connection, you click the “run test”
button to start your test, when the test is done, its results
are displayed on the program log as shown in Fig.6.
B. Experiment Management
In this tab, user can execute commands on remote nodes
or upload files to them. The basic functions that can be
performed within this tab are:
• Node selection: Nodes can be selected by clicking on
them in the node list table or by clicking on the
nodes' icons on the map. Information about a node
can be obtained by hovering the mouse pointer on
it, this will display a tool tip showing the node
details.
• Command execution: To execute a command, you
simply select the nodes on which you want to run
your command, insert the command text with all
variables in the command text box, and then click
the “run command” button. The output from each
node will be displayed on the program log.
• Uploading Files: To upload files to nodes, go to the
“copy file” tab, from there you can select the file
which you want to upload using the “browse”
button, you can also specify the destination
directory on the remote nodes which you want to
store your file in, after selecting the nodes you click
“upload file” to start uploading your file. When
uploading is done, the log will display the output
from the remote consoles to verify that the process
is successful, if there are no error messages then the
file is uploaded.
C. Connection Testing
© 2012 ACADEMY PUBLISHER
Figure 6: Adding nodes to a path, : The test results
showing on the program log
D. Illustrative Examples
We consider two scenarios to show the functionality and
flexibility that ConTest offers in setting up and executing
experiments over PlanetLab. In the first scenario, we
show how to use the tool to obtain information about the
disk usage percentage on each node within a slice. In the
second scenario, we show how to perform round trip time
measurements between two nodes and how to assign
roles for each node.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Example 1-Retrieving Disk Usage Percentage: In this
example, we show how to access multiple PlanetLab
nodes simultaneously and perform a simple test to obtain
the percentage of the disk space used on each node within
a slice. In order to perform this test, a shell script
responsible for obtaining the disk space usage has been
uploaded on each node within our slice. This was done by
selecting all nodes from the list in the node list view and
uploading the shell script “diskUsuage.sh”. To find the
disk usage percentage on a node, the shell script was
executed via the execute command tab where the path of
the script “testScript/deskUsage.sh” is specified. Disk
usage percentage is displayed in the program log area as
shown in Fig. 7.
Example 2: Round Trip Time (RTT) Measurements
In ConTest, the connection test interface allows users to
specify the role of each node in an experiment. In this
regard, a node can be specified to be a client, a server, or
a forwarding node (i.e., a proxy). Also, it allows them to
upload the code that is suitable for the role of each node.
Measuring the round trip time between two nodes is a
simple example that we describe here to illustrate this
functionality.
Fig. 8 shows the results of measuring RTT between a
client (the green node) and a server (the pink node). It
also shows the results of measuring RTT between the
same nodes assuming that the traffic is forced to pass
through another node (the proxy) represented by orange
in this example. Such setup is commonly used for indirect
communication and it can serve for testing triangular
Figure 7: Using ConTest to obtain disk usage percentage of each node within a slice (Example 1)
Figure 8. The results of measuring round trip time between two nodes (Example 2)
© 2012 ACADEMY PUBLISHER
169
170
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
routing or even studying some characteristics of fast flux
networks which is becoming a very important problem
[11][12].
It is important to emphasize here that using ConTest to
perform such experiments over PlanetLab provides great
deal of flexibility and saves a lot of time to setup and run
the experiment and collect the related results. So instead
of using the CLI to access each node individually to
upload the files, execute commands or setting up tests,
ConTest automates this process and saves the user's time
and effort.
V. RELATED WORK
Since the deployment of PlanetLab, there have been a lot
of efforts to simplify the setup and control of experiments
over this distributed testbed. These efforts focused mainly
on providing a graphical user interface to select nodes,
execute commands, and perform different measurement
tests. These tools typically take advantage of the
PlanetLab Central (PLC) which provides detailed
information about PlanetLab nodes such as host name,
geographical location, TCP connectivity. It also provides
convenient interfaces for adding/removing nodes. Planet
lab manager [13] is one of these tools. It basically allows
users to choose nodes for their slices and then to choose
some of these nodes for an experiment based on the
common information about these nodes. It has the
capability of deploying experiment files and execution
single/multiple command on every node in parallel. It
also provides means to monitor the progress of an
experiment and to view the output from the nodes.
PlanetLab application manager [14] is another tool that
was designed to simplify deployment and monitoring and
running applications over PlanetLab. It has several
features that enable users to centrally manage their
experiments and monitor their applications. The tool is
composed of two components: the server side which
requires access to aPostagreSQl/MySQL Database
element and a PhP web server to allow web access. On
the other hand, the client side is basically a shell script
that runs under bash. The client side shell scripts require
specific customizations making it a little bit complicated.
Stork is a software installation utility service that installs
and maintains software for other services [15]. It allows
users to install software on a large number of PlanetLab
nodes and keep that software updated for a long period of
time. It also has and a novel security system that
empowers users and package creators while reducing
trust in a repository administrator. As a results Stork
provides a secure, scalable, effective way to maintain
tarball and RPM packages and software on large
networks.
Other tools have been developed to simplify the process
of evaluating the characteristics of PlanetLab nodes, thus
allowing users to select suitable nodes for their
experiments. e. g., pShell [16], CoDeeN [17] [18]. The
pShell is a Linux shell that provides basic commands to
© 2012 ACADEMY PUBLISHER
interact with PlanetLab nodes. However, for pShell users,
they still need to manually select suitable nodes.
ConTest share many of these features with other tools.
The most obvious difference between ConTest and other
tools is that ConTest provide a visual way to select nodes
within a slice using the world map view. It also, allow
users to visually construct the network topology and
project it on the map. Moreover, users can specify the
role of each node to be client, server, or forwarding node.
VI. CONCLUSION AND FUTURE WORK
In this paper, we have discussed the design and
implementation of ConTest, a graphical user interfacebased tool that enables researchers to visualize
experiments over PlanetLab network. ConTest is a
powerful graphical tool that helps PlanetLab users access
nodes within their slices, upload files and execute
commands easily, with no complications and with much
fewer security restrictions compared to the command line
access, which is the default way of accessing PlanetLab
network. It also helps users perform different tests on the
nodes they choose and automates gathering and
summarizing the results for these tests. Also, it gives the
users the ability to utilize its connection automation to
run their own tests using their custom testing
applications. This tool would be of a great help So this
tool is a very promising tool as it has unlimited
improvement possibilities which we hope we will be able
to explore to make it a more effective tool in the world of
distributed applications testing and development.
As part of our ongoing efforts to improve ConTest’s
functionality and usability, we seek to achieve the
following goals:
• In the current implementation, the user can
browse remote machine folders using the
browsing commands (like ls). Future release of
ConTest will enable browsing directories on a
remote node graphically to increase the usability
and ease of use.
• Enhance our tool's testing capabilities by adding
some additional statistics. Users of the tool can
easily help in improving this aspect since the test
codes are separated from the main tool
functionality, so virtually any codes can be
tested and used.
• Improving the graphical design of the tool.
Specifically the map view. In particular, we will
increase the accuracy of the coordinate mapping
and add the zooming feature to make the map
interface easier to use.
• Even though we enable the user to choose the
test codes, it is still, somehow, a limited feature.
On the far view, we want to enable the users to
conduct their own experiments based on the
connection setup they choose, instead of limiting
them by simple connection testing codes.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
ACKNOWLEDGMENT
The authors would like to thank Dr. Mohammad
Fraiwan for his constructive thoughts and comments to
improve this work.
REFERENCES
[1]
[2]
PlanetLab. [Online] http://www.planetlab.org.
[18] M. J. Freedman, E. Freudenthal, & D. Mazires,
(2004) “Democratizing content publication with coral”,
In Proceedings of the 1st USENIX Symposium on
Networked Systems Design and Implementation (NSDI
’04).
[3] V. Ramasubramanian, & E.G. Sirer, (2004) “O(1) lookup
performance for power-law query distributions in peer-topeer overlays”, In NSDI , pp 99-112.
[4] K. Park, & V.S. Pai, (2006) “Scale and performance in
the coblitz large-file distribution service”, In Proc. 3rd
Symposium on Networked Systems Design and
Implementation (NSDI 06).
[5] N. Spring, D. Wetherall, & T. Anderson, (2002)
“Scriptroute: A public internet measurement facility”.
[6] Nokia
Corporation,
Qt
Creator,
qt.nokia.com/products/developer-tools, 2010.
[7] Qt Centre Forum - Qt Centre Community Portal,
www.qtcentre.org.
[8] Charles Morin, Randy, "How to PING", 2001.
[9] Forouzan and Behrouz, "Data Communications And
Networking", 2007.
[10] Iperf project, sourceforge.net/projects/iperf, 2008.
[11] Holz, T., Gorecki, C., Rieck, K., Freiling, F.: Measuring
and detecting fast-flux service networks. In: Proceedings
of the Network & Distributed System Security
Symposium (2008).
[12] Passerini, E., Paleari, R., Martignoni, L., Bruschi, D.:
FluxOR: detecting and monitoring fast-flux service
networks. Detection of Intrusions and Malware, and
Vulnerability Assessment pp. 186–206 (2008).
[13] PlanetLab
Experiment
Manager,
http://www.cs.washington.edu/research/networking/cplan
e/
[14] R.
Huebsch.
Planetlab
application
manager.
http://appmanager.berkeley.intel-research.net/, 2004.
[15] J. Cappos and J. Hartman. Why it is hard to build a longrunning service on PlanetLab. In Proc. Workshop on Real,
Large Distributed Systems (WORLDS), San Francisco,
CA, Dec. 2005.
[16] B. Maniymaran, pShell: An Interactive Shell for
Managing Planetlab Slices, http://cgi.cs.mcgill.ca/
anrl/projects/pShell/index.php/
[17] L.Wang, V. Pai and L. Peterson, The Effectiveness of
Request Redirection on CDN Robustness, Proceedings of
the 5th OSDI Symposium, December 2002.
[18] CoDeeN,
Princeton
University:
http://codeen.cs.princeton.edu/
© 2012 ACADEMY PUBLISHER
171
Basheer Al-Duwairi received the PhD
and MS degrees in computer
engineering from Iowa State University
in Spring 2005 and Spring 2002,
respectively. Prior to this, he received
the BS degree in electrical and
computer engineering from Jordan
University of Science and Technology
(JUST) Irbid, Jordan in 1999. He has
been an assistant professor at the Department of Network
Engineering and Security at JUST since fall 2009; prior to this,
he was an assistant professor in the Computer Engineering
Department at the same university from fall 2005 to fall 2009.
His research expertise are in the areas of trusted Internet
encompassing Internet Infrastructure Security focusing on
DDoS, Botnets, and P2P security, Wireless Network security,
and Resource Management in real-time systems. He has
coauthored several research papers in these fields. He has given
tutorials at reputed conferences, served as a member of
technical program committee and session chair in many
conferences. He also served as Assistant Dean for Student
Affairs, Chairman of Computer Engineering Department, Vice
Dean for Computer and Information Technology, and the
Director of Computer and Information Cenetr, all at JUST.
http://www.just.edu.jo/~basheer
Mohammad Marei, Malek Ireksossi, and Belal Rehani are
undergraduate students at the department of computer
engineering at Jordan University of Science & Technology.
They are expected to graduate in summer 2011.
172
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Context-oriented Software Development
Basel Magableh
Trinity College Dublin, Ireland, Ireland
Email: magablb@cs.tcd.ie
Stephen Barrett
Trinity College Dublin, Ireland, Ireland
Email: Stephen.barrett@tcd.ie
Abstract— Context-oriented programming is an emerging
technique that enables dynamic behaviour variation based
on context changes. In COP, context can be handled directly
at the code level by enriching the business logic of the
application with code fragments responsible for performing
context manipulation, thus providing the application code
with the required adaptive behavior. Unfortunately, the
whole set of sensors, effectors, and adaptation processes is
mixed with the application code, which often leads to poor
scalability and maintainability. In addition, the developers
have to surround all probable behavior inside the source
code. As an outcome, the anticipated adjustment is
restricted to the amount of code stubs on hand offered by
the creators. Context-driven adaptation requires dynamic
composition of context-dependent parts. This can be
achieved trough the support of a component model that
encapsulates the context-dependent functionality and
decouples them from the application’s core-functionality.
The complexity behind modeling the context-dependent
functionality lies in the fact that they can occur separately
or in any combination, and cannot be encapsulated because
of their impact across all the software modules. Before
encapsulating crosscutting context-dependent functionality
into a software module, the developers must first identify
them in the requirements documents. This requires a formal
development paradigm for analyzing the context-dependent
functionality; and a component model, which modularizes
their concerns. COCA-MDA is proposed in this article as
model driven architecture for constructing self-adaptive
application from a context oriented component model.
Index Terms— Adaptable middleware, Context oriented
component, Self-adaptive application, Object.
I. INTRODUCTION
There is a growing demand for developing applications
with aspects such as context awareness and self-adaptive
behaviors. Self-adaptive software evaluates its own
behavior and changes its behavior when the evaluation
indicates that it is not accomplishing what the software is
intended to do, or when better functionality or
performance is possible. Traditionally, self-adaptability is
needed to handle complexity, robustness of unanticipated
conditions, changing of priorities and polices governing
the objective goals, and changing conditions in the
contextual environment. Hirschfeld et al. [1] considered
context to be any information that is computationally
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.172-180
accessible and upon which behavioral variations depend.
A context-dependent application adjusts its behavior
according to context conditions arising during execution.
The techniques that enable applications to handle the
contextual application are generally known as “contexthandling” techniques. Context handling is of vital
importance for developers and for self-adaptive
architecture since it provides dynamic behavioral
adaptation and makes it possible to produce more useful
computational services for end users in the mobile
computing environment [2]. The mobile computing
environment is heterogeneous and dynamic. Everything
from devices used, resources available, network
bandwidths, to user context, can change drastically at
runtime [3]. This presents the application developers with
the challenge of tailoring behavioral variations to each
specific user and context. With the capacity to move and
the desire to be socially collaborative, mobile computing
users might benefit from the self-adaptability and the
context-awareness features that are supported by selfadaptive applications.
This article focuses on describing a development
paradigm for Context-Oriented Programming, which
enables self-adaptability features in this emerging class of
applications. The development methodology Context
Oriented Component-based Application Model-Driven
Architecture
(COCA-MDA)
modularizes
the
application’s context-dependent behavior into contextoriented components. The components separate the
application’s functional concerns from the extrafunctional concerns. The application is organized into
two casually connected layers: the base layer, which
provides the application’s core structure, and the metalayer, where the COCA-components are located, and
which provides compassable units of behavior. The
component model design (COCA-components) has been
proposed in previous work in [4]. A COCA-component
refers to any subpart of the software system that
encapsulates a specific context-dependent functionality in
a unit of behavior contractually specified by interfaces
and explicit dependences. The result from this
methodology is a component-based application described
by an architecture description language, COCA-ADL.
COCA-ADL is used to bridge the gap between the
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
software models in the platform-independent model of
the MDA and the software architecture runtime model.
Such employment of the ADL decouples the application’s
architecture design from the target platform
implementation.
The rest of the article is structured as follows. Section
II discusses behavioral variability support in ContextOriented Programming and aspects. Section III describes
the rationale for providing a development paradigm for
context-oriented Programming. Section IV describes the
COCA-component model. Section V describes the
COCA-ADL elements. The COCA-MDA phases are
described in Section VII. Section VIII demonstrates a
case study designed using the COCA-MDA and
implemented with the COCA-middleware.
II.
VARIABILITY MANAGEMENT WITH CONTEXTORIENTED PROGRAMMING AND ASPECTS
Compositional adaptation enables an application to
adopt a new structure/behavior for anticipating concerns
that were unforeseen during the original design and
construction. Normally, compositional adaptation can be
achieved using the separation of concerns techniques,
computational reflection, component-based design, and
adaptive middleware [5]. The separation of concerns
enables the software developers to separate the functional
behavior and the crosscutting concerns of the selfadaptive applications. The functional behavior refers to
the business logic of an application [5]. Context-driven
behavioral variations are heterogeneous crosscutting
concerns and a set of collaborated aspects that extend the
application behavior in several parts of the program and
have an impact across the whole system. Such behavior is
called crosscutting concerns. Crosscutting concerns are
properties or areas of interest such as quality of service,
energy consumption, location awareness, users’
preferences, and security. This work considers the
functional behavior of an application as the basecomponent that provides the user with context-free
functionality. On the other hand, context-dependent
behavior variations are considered as crosscutting
concerns that span the software modules in several
places.
Context-oriented programming is an emerging
technique that enables context-dependent adaptation and
dynamic behavior variations [6, 7]. In COP, context can
be handled directly at the code level by enriching the
business logic of the application with code fragments
responsible for performing context manipulation, thus
providing the application code with the required adaptive
behavior [8]. Unfortunately, the whole set of sensors,
effectors, and adaptation processes is mixed with the
application code, which often leads to poor scalability
and maintainability [9]. In general, the proposed COP
approaches support fine-grained adaptation among the
variant behaviour that were introduced at the compile
time, A special compiler is needed for performing the
context handling operation. To best of our knowledge,
COP does not support dynamic composition of software
© 2012 ACADEMY PUBLISHER
173
modules and have no support for introducing new
behaviour/or adjusting the application structure to
anticipate the context changes. In addition, the developers
have to surround all probable behavior inside the source
code. As an outcome, the anticipated adjustment is
restricted to the amount of code stubs on hand offered by
the creators. On the other hand, it is impractical to
forecast all likely behaviors and program them at the
source code.
For a more complex context-aware system, the same
context information would be triggered in different parts
of an application and would trigger the invocation of
additional behavior. In this way, context handling
becomes a concern that spans several application units,
essentially crosscutting into the main application
execution. A programming paradigm aiming at handling
such crosscutting concerns (referred to as aspects) is
aspect-oriented programming (AOP) [10]. In contrast to
COP, Using the AOP paradigm, context information can
be handled through aspects that interrupt the main
application execution. In order to achieve self-adaptation
to context in a manner similar to COP, the contextdependent behavioral variations must be addressed.
Unfortunately,
the
aspect-oriented
development
methodology can be used to handle homogeneous
behavioral variations where the same piece of code can
be invoked in several software modules [11, 12], and it
does not support adaptation of aspects to context in what
is called context-driven adaptation [9]. Moreover, static
AOP is classified as a compositional adaptation
performed in compile time [5]; anticipating context
changes at runtime is not an option, especially with the
presence of unforeseen changes. Another approach
supported by AOP is called dynamic weaving for aspects
[13]; this injects the code in the program execution
whenever a new behavior is needed. However, existing
AOP languages tend to add a substantial overhead in both
execution time and code size, which restricts their
practicality for small devices with limited resources [14].
III.
RATIONALE
Context changes are the causes of adaptation. A
context-driven adaptation requires the self-adaptive
software to anticipate its context-dependent variations.
The context-dependent variation can be classified into
actor-dependent, system-dependent, and environmentdependent behavior variations. The complexity behind
modeling these behavior variations lies in the fact that
they can occur separately or in any combination, and
cannot be encapsulated because of their impact across all
the software modules. Context-dependent variations can
be seen as collaboration of individual features (aspects)
expressed in requirements, design, and implementation,
and are sufficient to qualify as heterogeneous crosscutting
concerns in the sense that different code fragments are
applied to different program parts. Before encapsulating
crosscutting context-dependent behaviors into a software
module, the developers must first identify them in the
requirements documents. This is difficult to achieve
174
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
because, by their nature, context-dependent behaviors are
tangled with other behaviors, and are likely to be
included in multiple parts (scattered) of the software
modules. Using intuition or even domain knowledge is
not necessarily sufficient for identifying the contextdependent parts of self-adaptive applications. This
requires a formal procedure for analyzing them in the
software requirements and separating their concerns.
Moreover, a formal procedure for modeling these
variations is needed. Such analysis and modeling
procedures can reduce the complexity in modeling selfadaptive applications. In this sense, a formal development
methodology can facilitate the development process and
provide new modularization of a self-adaptive software
system in order to isolate the context-dependent from the
context-free functionalities. Such a methodology, it is
argued, can decompose the software system into several
behavioral parts that can be used dynamically to modify
the application behavior based on the execution context.
Behavioral decomposition of a context-aware
application can provide a flexible mechanism for
modularizing the application into several units of
behavior. Because each behavior realizes a specific
context condition at runtime, such a mechanism requires
separation of the concerns of context handling from the
concern of the application business logic. In addition,
separation of the application’s context-dependent and
context-independent parts can support a behavioral
modularization of the application, which simplifies the
selection of the appropriate parts to be invoked in the
execution whenever a specific context condition is
captured. The adaptive software operates through a series
of substates (modes). The substates are represented by j,
and j might represent a known or unknown conditional
state. Examples of known states in the generic form
include detecting context changes in a reactive or
proactive manner.
a set of context-driven and context-free states. At
runtime, the middleware transforms the self-adaptive
software form statei into statei+1, considering a specific
context condition tjk, as shown in Figure 1. This enables
the developer to clearly decide which part of the
architecture should respond to the context changes tjk,
and provides the middleware with sufficient information
to consider a subset of the architecture during the
adaptation; this enhances the adaptation process, impact,
and cost and reduces the computation overhead from
implementing this class of applications in mobile devices.
Self-adaptive and context-aware applications can be
seen as the collaboration of individual behavioral
variations expressed in requirements, design, and
implementation. This article contributes by proposing a
model-driven architecture (COCA-MDA) integrated with
a behavioral decomposition technique, based on
observation of the context information in requirements
and modeling. As a result of combining a decomposition
mechanism with MDA, a set of behavioral units is
produced. Each unit implements several contextdependent functionalities. This requires a component
model that encapsulates these code fragments in distinct
architecture units and decouples them from the corefunctionality components. This is what motivates the
research towards proposing a context-oriented component
model (COCA-component). Context-driven adaptation
requires dynamic composition of context-dependent
parts, which enables the developer to add, remove, or
reconfigure components within an application at runtime.
Each COCA-component embeds a specific contextdependent functionality (C+S)ji, realized by a contextoriented component (COCA-component) model. Each
COCA-component realizes several layers that encapsulate
a fragment of code related to a specific software mode
layer(Mjk), as shown in Figure 1. The developers have
the ability to provide a decision policy(jk) for each
decision point (j) whenever a specific context-related
condition is captured. Hereafter, the COCA-components
are dynamically managed by COCA-middleware and
their internal parts to modify the application behavior.
The COCA-middleware performs context monitoring,
dynamic decision-making, and adaptation, based on
policy evaluation. The decision policy framework is
maintained in modeling and runtime time.
Figure 1: Behavioral Decomposition Model
In the presence of uncertainty and unforeseen context
changes, the self-adaptive application might be notified
about an unknown condition prior to the software design.
Such adaptation is reflected in a series of context-system
states. (C+S)ji denotes the ith combination of contextdependent behavior, which is related to the decision
points j by the notion mode Mjk. In this way, the
development methodology decomposes the software into
© 2012 ACADEMY PUBLISHER
IV.
CONTEXT-ORIENTED COMPONENT MODEL
(COCA-COMPONENT)
The COCA-component model was proposed in [21],
based on the concept of a primitive component
introduced by Khattak et al. in [17] and Context-Oriented
Programming (COP) [13]. COP provides several features
that fit the requirements of a context-aware application,
such as behavioral composition, dynamic layers
activation, and scoping. This component model
dynamically composes adaptable context-dependent
applications based on a specific context-dependent
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
functionality. The authors developed the component
model by designing components as compositions of
behaviors, embedding decision points in the component
at design time to determine the component behaviors, and
supporting reconfiguration of decision policies at runtime
to adapt behaviors.
175
At this stage, each COCA-component must adopt the
COCA-component model design. A sample COCAcomponent is shown in Figure 2; it is modeled as a
control class with the required attributes and operations.
Each layer entity must implement two methods that
collaborate with the context manager. Two methods
inside the layer class, namely ContextEntityDidChanged
and ContextEntityWillChanged, are called when the
context manager posts the notifications in the form
[NotificationCeneter
Post:ContextConditionDidChanged]. This triggers the
class
layer
to
invoke
its
method
ContextEntityDidchanged, which embeds a subdivision
of the COCA-component implementation.
V. COCA-ADL: A CONTEXT-ORIENTED
COMPONENT-BASED APPLICATION ADL
Figure 2: COCA-component Conceptual Diagram
The COCA-component has three major parts: a static
part, a dynamic part, and ports. The component itself
provides information about its implementation to the
middleware. The COCA-component has the following
attributes: ID, name, context entity, creation time,
location, and remote variable. The Boolean attribute
remote indicates whether or not the components are
located on the distributed environment. The decision
policy and decision points are attributes with getter and
setter methods. These methods are used by the
middleware to read the attached PolicyID and manipulate
the application behavior by manipulating the decision
policy.
The COCA-component handles the implementation of
a context-dependent functionality through employing the
delegate design pattern [6], so the adaptation manager
invokes these components whenever the COCAcomponent is notified by the context manager. A delegate
is a component that is given an opportunity to react to
changes in another component or influence the behavior
of another component. The basic idea is that two
components coordinate to solve a problem. A COCAcomponent is very general and intended for reuse in a
wide variety of contextual situations. The basecomponent stores a reference to another component, i.e.,
its delegate, and sends messages to the delegate at critical
times. The messages may only inform the delegate that
something has happened, giving the delegate an
opportunity to do extra processing, or the messages may
ask the delegate for critical information that will control
what happens. The delegate is typically a unique custom
object within the controller subsystem of an application
[6].
© 2012 ACADEMY PUBLISHER
The aim of this section is to introduce the architecture
description language COCA-ADL. COCA-ADL is an
XML-based language used to describe the architecture
produced by the development methodology COCAMDA. COCA-ADL is used to bridge the gap between the
application models and the implementation language.
Thus, it enables the architecture to be implemented by
several programming languages.
Figure 3: COCA-ADL Elements
COCA-ADL is designed as a three-tier system. The
first level consists of the building blocks, i.e., the
components, including the COCA-component and basecomponent. The second refers to connectors, and the third
refers to the architecture configuration, which includes a
full description of the state-machine models, which
describes an activity diagram plus the decision policies’
syntax. Figure 3 shows the main elements of COCAADL. Each element is associated with an architecture
template type. The main features provided by the element
types are instantiation, evolution, and inheritance.
VI. COCA-MDA DEVELOPMENT APPROACH
COCA-MDA has adopted the component collaboration
architecture (CCA) and the entity model. The CCA
details how to model the structure and behavior of the
components that comprise a system at varying and mixed
levels of granularity. The entity model describes a metamodel that may be used to model entity objects that are a
representation of concepts in the application problem
domain and define them as composable components [17].
176
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
COCA-MDA partitioning the software into three
viewpoints: the structure, behavior, and enterprise
viewpoints. The structure viewpoint focuses on the core
component of the self-adaptive application and hides the
context-driven component. The behavior viewpoint
focuses on modeling the context-driven behavior of the
component, which may be invoked in the application
execution at runtime. The enterprise viewpoint focuses on
remote components or services, which may be invoked
from the distributed environment. The design of a
context-aware application according to the COCA-MDA
approach was proposed in [18]. The use of COCA-MDA
for developing self-adaptive applications for indoor
wayfinding for individuals with cognitive impairments
was proposed in [19]. Evaluating the COCA-MDA
productivity among the development effort was
demonstrated in [20]. This article focuses on describing
in detail the process of analyzing and classifying the
software requirements, and how the software is designed
through the platform-independent model and the
platform-specific model. Model transformation and code
generation were discussed in a previous work [18].
behavioral models, and the context
transformed into the COCA-ADL model.
model
are
A. Computational Independent Model
In the analysis phase, the developers analyze several
requirements using separation of concern techniques. The
developers focus on separating the functional
requirements from the extra-functional requirements as
the first stage. They then separate the user and context
requirements from each other. There are two subtasks in
the analysis phase.
VII. CONTEXT-ORIENTED COMPONENT-BASED
APPLICATION EXAMPLE
IPetra is a tourist-guide application that helps the client
to determine the bravura historical city of Petra, Jordan.
IPetra offers a map–client interface maintained by an
augmented reality browser (ARB). The browser exhibits
many points of interest (POI) inside the physical outlook
of the tool’s camera. Information related to every POI is
exhibited inside the camera overlay outlook. The POIs
comprise edifices, tourist services sites, restaurants,
hotels, and ATMs in Petra. The AR browser offers an
instantaneous live direct physical display inside the
portable camera. When the client positions the portable
camera in the direction of a building, an explanation
confined to a small area related to that edifice is shown to
the client.
Constant use of the device’s camera, backed by
attaining data from many sensors, can consume the tool’s
resources. This needs the application to adjust its tasks
among several contexts to maintain quality of service
without disrupting the function’s tasks. The function
requires frequent updates of client position, network
bandwidth, and battery level.
Figure 6 summarizes the modeling tasks, using the
associated UML diagrams. The developer starts analysis
of an application scenario to capture the requirements.
The requirements are combined in one model in the
requirements diagram. The requirements diagram is
modeled using a use-case diagram that describes the
interaction between the software system and the context
entity. The use-case is partitioned into two separate
views. The core-structure view describes the core
functionality of the application. The extra-functionality
object diagram describes the COCA-component
interaction with the core application classes. The state
diagram and the activity diagram are extracted from the
behavioral view. Finally, the core structure, the
© 2012 ACADEMY PUBLISHER
Figure 6: Modelling tasks
1) Task 1: Requirements capturing by textual analysis:
In this task, the developer identifies the candidate
requirements for the illustration scenario using a
textual analysis of the application scenario. It is
recommended that the developer identifies the
candidate actors, use-cases, classes, and activities, as
well as capturing the requirements in this task. This can
be achieved by creating a table that lists the results of
the analysis. This table provides an example of a data
dictionary for the proposed scenario.
2) Task2: Identifying the extra-functional requirements
and relating them into the middleware functionality:
The first step in the process is to understand the
application’s execution environment. The context is
classified in the requirements diagram, based on its
type, and whether it comes from a context provider or
consumer. A context can be generated from a physical
or logical source (i.e., available memory), or resources
(i.e., battery power level and bandwidth). The
representations of sensors and resources on the
application that is going to consume them at runtime
refers to the context consumers.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Figure 7: Requirements UML profile
The next level of requirements classification is to
classify the requirements based on their anticipation
level; this can be foreseeable, foreseen, or unforeseen.
This classification allows the developer to model the
application behavior as much as possible and to plan for
the adaptation actions. However, to facilitate this
classification framework, a UML profile is designed to
support the requirements analysis and to be used by the
software designer, as shown in Figure 7.
As shown in Figure 8, extra-functional requirements
are captured during this task, for example, requirement
number 3: adapt the location service. IPetra is required to
adapt its behavior and increase the battery life. This is
achieved by adopting a location service that consumes
less power. For example, if the battery level is low, the
IPetra application switches off the GPS location services
and uses the cell-tower location services. Using an IPbased location reduces the accuracy of the location but
saves battery energy. In addition, the application may
reduce the number of POIs it displays to the most recent
device location. Moreover, the application reduces the
frequency of the location updates. On the other hand, if
the battery level is high and healthy, IPetra uses the GPS
service with more accurate locations. The application
starts listening for all events in the monitored region
inside Petra city.
3) Task 3: Capturing user requirements: This task
focuses on capturing the user’s requirements as a
subset of the functional requirements, as shown in the
UML profile in Figure 7. This task is similar to a
classical requirement-engineering process where the
developers analyze the main functions of the
application that achieve specific goals or objectives.
B. Modelling: Platform Independent Model
In order to be aware of possible resources and context
variations and the necessary adaptation actions, a clear
analysis of the context environment is the key to building
dynamic context-aware applications.
4) Task 4: Resources and context entity model
Resources and context Model refers to generic a
overview of the underlying device’s resources, sensors,
and logical context provider. This diagram is modelling
the engagement between the resources and the application
under development. It facilitates the developer to
understand the relationship between the context providers
and their dependency.
© 2012 ACADEMY PUBLISHER
177
5) Task 5: use-cases
The requirements diagram in Figure 8 represents the main
inputs for this task. Each requirement is incorporated into
a use-case, and the developers identify the actor of the
requirement. An actor could be a user, system, or
environment. The use-cases are classified into two
distinct classes, i.e., the core functionality and extended
use-cases, by the context conditions. The first step is to
identify the interaction between the actor and the
software functions to satisfy the user requirement in a
context-free fashion. For example, the displaying POIs
functionality in the figure is context independent in the
sense that the application must provide it, regardless of
the context conditions. All these use-cases are modeled
separately, using a class diagram that describes the
application core-structure or the base-component model,
as shown in the following task.
Figure 8: Functional and extra-functional partial requirements diagram
6) Task 6: modeling the application core-structure
In this task, a classical class diagram models the
components that provide the application’s core functions.
These functions are identified from the use-case diagram
in the previous task. However, the class diagram is
modeled independently from the variations in the context
information. For this scenario, some classes, such as
“Displaying POIs”, “Route-planningUI”, “CameraUI”,
“MapUI”, and “User Interface”, are classified to be on the
application core. These classes provide the core functions
for the user during his tour of Petra city. Figure 9 shows
the core-structure class-model without any interaction
with the context environment or the middleware.
7) Task 7: identifying application-variant behavior
(behavior view):
In this phase, the developers specify how the application
can adapt to context conditions to achieve a specific goal
or objectives. After specifying the core elements of the
application in the previous task, the behavioral view is
identified in this task. This task identifies when and
where an extra-functionality can be invoked in the
application execution. This means the developer has to
analyze the components involved, their communication,
and possible variations in their subdivisions, where each
178
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
division realizes a specific implementation of that
COCA-component.
To achieve this integration, the developers have to
consider two aspects of the context-manager design: how
to notify the adaptation manager about the context
changes, and the how the component manager can
identify the parts of the architecture that have to respond
to these changes. These aspects can be achieved by
adopting the notification design pattern in modeling the
relation between the context entity and the behavioral
component. Hereafter, these extra-functionalities are
called the COCA-components. Each component must be
designed on the basis of the component model described
in Figure 2.
Figure 9: I-Petra Core-Classes structure
The IPetra application is modularized into several
COCA-components. Each component models one extrafunctionality such as the LocationCOCA−component in
Figure 10. The COCA-component sublayers implement
several context-dependent functionalities that use the
location service. Each layer is activated by the
middleware, based on context changes. After applying the
observer design pattern and the COCA-component model
to the use-cases, the class diagram for the middleware
functionality “Update Location” can be modeled as
shown in Figure 10. Figure 10 shows a COCAcomponent modelled to anticipate the ’direction output’.
The COCA-component implements a delegate objects
and sub layers; each layer implements a specific contextdependent function. The COCA-middleware uses this
delegate object to redirect the execution among the sub
layers, based on the context condition.
Invoking different variations of the COCA-component
requires identification of the application architecture,
behavior, and the decision policies in use. As mentioned
before, these decision policies play an important role in
the middleware functions, which use them in handling,
the architecture evolution, and the adaptation action. The
model in Figure 10 helps the developer to extract the
decision policies and the decision points from the
interactions between the context entities and the COCAcomponents. Decision policies are used by the
middleware to select suitable adaptation actions among
specific context conditions.
The application behavioural model is used to
demonstrate the decision points in the execution that
© 2012 ACADEMY PUBLISHER
might be reached whenever internal or external variables
are found. This decision point requires several parameter
inputs to make the correct choice at this critical time.
Using the activity diagram, the developers can extract
numerous decision polices. Each policy must be modelled
in a state diagram, for example, the Policy: Camera
Flashes is attached to the ’Camera flashes’ COCAcomponent. The policy syntax can be described by the
code shown in listing 1.
The IPetra application has been implemented in two
distinct versions, i.e. with and without the COCAmiddleware. The Instruments is a tool application for
dynamically tracing and profiling IPhone devices. The
battery life has been measured by Energy Diagnostics
Instruments running on the device [22]. The energy
Diagnostic used to measure the battery while the device is
not connected to external power supply. The experiments
show that the COCA IPetra application saved the batteryconsuming level by 13% in addition to its self-tuning
adaptability. Fig. 14 shows the experimental results for
energy usage. The IPetra implementation without
adopting the COCA-platform consumes more energy
during context monitoring, draining the battery faster. On
the other hand, when the same application adopts the
COCA-middleware, the application is able to adapt its
behaviour and use less energy. The adaptation time for
handling low and high battery-levels are shown in Figure
13. It is worth mentioning here that when the battery level
is low, the COCA-middleware allocates less memory
because of the size of the COCA-component, which is
small compared to its implementation.
Figure 10: Extra-functionality Object Diagram of the Context Oriented
Components
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
If ( direction is Provided && Available memory >=
50
&& CPU throughput <= 89 && light level >= 50
&& BatteryLevel >= 50) then {EnableFlashes();}
Else If ( BatteryLevel < 50 || LightLevel < 50 )
then {DisableFlashes(); SearchForPhotos();}
else If( BatteryLevel < 20)
then DisableFlashes();
Listing 1: Adaptation time (ms) and memory allocation (KB)
to anticipate the evolution of both functional and nonfunctional requirements. The decision policies require
more development with respect to policy mismatch and
resolution. This is in line with an improvement in terms
of self-assurance and dynamic evaluation of the
adaptation output.
REFERENCES
[1]
[2]
[3]
Figure 13: Adaptation time (ms) and memory allocation (KB)
[4]
[5]
[6]
Figure 14: Energy usage for IPetra application.
[7]
VIII. EXPERIMENTS EVALUATION
IX. CONCLUSIONS AND FUTURE WORKS
This article described a development paradigm for
building context-oriented applications using a
combination of Model-Driven Architecture that generates
an ADL, which presents the architecture as a
components-based system, and a runtime infrastructure
(middleware) that enables transparent self-adaptation
with the underlying context environment.
Specifically, a Model-Driven Architecture is used to
demonstrate a new approach to building context-aware
and self-adaptive applications by adopting a ModelDriven Architecture (COCA-MDA). COCA-MDA
enables developers to modularize applications based on
their context-dependent behaviors, enables developers to
separate context-dependent functionalities from the
application’s generic functionality, and enables dynamic
context-driven adaptation without overwhelming the
quality attributes.
The COCA-MDA needs to be improved with respect
to support for both requirement reflection and modeling
requirements as runtime entities. The requirement
reflection mechanism requires support at the modeling
level and at the architecture level. Reflection can be used
© 2012 ACADEMY PUBLISHER
179
[8]
[9]
[10]
[11]
[12]
[13]
R. Hirschfeld, P. Costanza, and O. Nierstrasz, “Contextoriented programming,” Journal of Object Technology,
vol. 7, no. 3, pp. 125–151, March 2008.
A. K. Dey, “Providing architectural support for building
context-aware applications,” Ph.D. dissertation, Georgia
Institute of Technology, Atlanta, GA, USA, 2000.
N. M. Belaramani, C.-L. Wang, and F. C. M. Lau,
“Dynamic component composition for functionality
adaptation in pervasive environments,” in Proceedings of
the The Ninth IEEE Workshop on Future Trends of
Distributed Computing Systems, ser. FTDCS ’03. plus
0.5em minus 0.4emIEEE Computer Society, 2003, pp.
226–232.
B. Magableh and S. Barrett, “Pcoms: A component model
for building context-dependent applications,” in
Proceedings of the 2009 Computation World: Future
Computing, Service Computation, Cognitive, Adaptive,
Content, Patterns, ser. COMPUTATIONWORLD ’09.
plus 0.5em minus 0.4emWashington, DC, USA: IEEE
Computer Society, 2009, pp. 44–48.
P. K. McKinley, S. M. Sadjadi, E. P. Kasten, and B. H. C.
Cheng, “Composing adaptive software,” Computer,
vol. 37, pp. 56–64, July 2004.
M. Gassanenko, “Context-oriented programming,” in
Proceedings of the EuroFORTH’93 conference,
Marianske Lazne (Marienbad), Czech Republic, 15-18
October 1998, pp. 1–14.
R. Keays and A. Rakotonirainy, “Context-oriented
programming,” in Proceedings of the 3rd ACM
international workshop on Data engineering for wireless
and mobile access, ser. MobiDe ’03, San Diego, CA,
USA, 2003, pp. 9–16.
M. Salehie and L. Tahvildari, “Self-adaptive software:
Landscape and research challenges,” ACM Trans. Auton.
Adapt. Syst., vol. 4, pp. 14:1–14:42, May 2009.
G. Kapitsaki, G. Prezerakos, N. Tselikas, and I. Venieris,
“Context-aware service engineering: A survey,” Journal
of Systems and Software, vol. 82, no. 8, pp. 1285–1297,
2009.
G. Kiczales, J. Lamping, A. Mendhekar, C. Maeda,
C. Lopes, J.-M. Loingtier, and J. Irwin, “Aspect-oriented
programming,” in ECOOP’97 âEuropean Conference on
Object-Oriented Programming, ser. Lecture Notes in
Computer Science. plus 0.5em minus 0.4emSpringer
Berlin / Heidelberg, 1997, vol. 1241, pp. 220–242.
S. Apel, T. Leich, and G. Saake, “Aspectual mixin layers:
aspects and features in concert,” in Proceedings of the
28th international conference on Software engineering,
ser. ICSE ’06. plus 0.5em minus 0.4emShanghai, China:
ACM, 2006, pp. 122–131.
M. Mezini and K. Ostermann, “Variability management
with feature-oriented programming and aspects,”
SIGSOFT Softw. Eng. Notes, vol. 29, pp. 127–136,
October 2004.
A. Popovici, T. Gross, and G. Alonso, “Dynamic weaving
for aspect-oriented programming,” in Proceedings of the
1st international conference on Aspect-oriented software
180
[14]
[15]
[16]
[17]
[18]
[19]
[20]
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
development, ser. AOSD ’02. plus 0.5em minus
0.4emNew York, NY, USA: ACM, 2002, pp. 141–147.
C. Hundt, D. Stöhr, and S. Glesner, “Optimizing aspectoriented mechanisms for embedded applications,” Lecture
Notes in Computer Science (including subseries Lecture
Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), vol. 6141 LNCS, pp. 137–153, 2010.
Y. Khattak and S. Barrett, “Primitive components:
towards more flexible black box aop,” in Proceedings of
the 1st International Workshop on Context-Aware
Middleware and Services: affiliated with the 4th
International Conference on Communication System
Software and Middleware (COMSWARE 2009), ser.
CAMS ’09. plus 0.5em minus 0.4emNew York, NY,
USA: ACM, 2009, pp. 24–30.
E. Buck and D. Yacktman, Cocoa design patterns,
2nd ed. plus 0.5em minus 0.4emDeveloper’s Library,
2010.
“Enterprise
collaboration
architecture
(eca)
specification,” http://www.omg.org/, pp. 1–202, Feb
2004.
B. Magableh and S. Barrett, “Objective-cop: Objective
context oriented programming,” in International
Conference on Information and Communication Systems,
ser. ICICS 2011, vol. 1, May 2011, pp. 45–49.
------, “Self-adaptive application for indoor wayfinding
for individuals with cognitive impairments,” in The 24th
IEEE International Symposium on Computer-Based
Medical Systems, ser. CBMS 2011, vol. In press,
Lancaster, UK, June 2011, pp. 45–49.
B. Magableh, “Model-Driven productivity evaluation for
self-adaptive Context-Oriented software development,” in
5th International Conference and Exhibition on Next
© 2012 ACADEMY PUBLISHER
Generation Mobile Applications, Services, and
Technologies (NGMAST’11), vol. In press, Cardiff,
Wales, United Kingdom, Sep. 2011.
[21] R. Anthony, D. Chen, M. Pelc, M. Perssonn, and
M. Torngren, “Context-aware adaptation in dyscas,” in
Proceedings of the Context-aware Adaptation
Mechanisms for Pervasive and Ubiquitous Services
(CAMPUS 2009), 2009, p. 15.
[22]
Ios
4.0
apple
developer
library.
http://developer.apple.com/library/ios/navigation/ (2010),
"[Online; accessed 1-April-2011]"
Basel Magableh received his MS degree in computer science
from New York Institute of Technology, NY, USA, in 2004. He
is currently a Ph.D. candidate at Distributed Systems Group,
Trinity College Dublin, Ireland. His research focuses in
integrating Model Driven Architecture with a component-based
system to construct self-adaptive and context-aware
applications.
He is a full-time lecturer in Grafton College of Management
Science, Dublin, Ireland. He was member of staff in the
National Digital Research Center of Ireland from 2008- 2011.
Stephen Barrett is currently a lecturer at Distributed Systems
Group, Trinity College Dublin, Ireland. His research centers on
middleware support for adaptive computing. (with particular
focus on model driven paradigms) and on large scale
applications research (particularly in the context of web search,
trust computation and peer and cloud computing) .
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
181
Aggregated Search in XML Documents
Fatma Zohra Bessai-Mechmache
Research Centre on Scientific and Technical Information, CERIST, Algiers, Algeria
zbessai@cerist.dz
Zaia Alimazighi
University of Science and Technology, USTHB, LSI, Algiers, Algeria
alimazighi@wissal.dz
Abstract—In this paper, we are interested in aggregated
search in structured XML documents. We present a
structured information retrieval model based on possibilistic
networks. Relations terms-elements and elements-document
are modeled through possibility and necessity. In this model,
the user’s query starts a process of propagation to recover
the elements. Thus, instead of retrieving a list of elements
that are likely to answer partially the user’s query, our
objective is to build a virtual elements that contain relevant
and non-redundant elements, that are likely to answer
better the query that elements taken separately. Indeed, the
possibilistic network structure provides a natural
representation of links between a document, its elements
and its content, and allows an automatic selection of
relevant and non-redundant elements. We evaluated our
approach using a sub-collection of INEX (INitiative for the
Evaluation of XML retrieval) and presented some results
for evaluating the impact of the aggregation approach.
possibilitic theory, it makes it possible to quantify in a
qualitative and quantitative way the various subjacent
links. it allows to express the fact that a term is certainly
or possibly relevant with respect to an element and/or a
document and to measure at which point an element (or a
set of elements) can necessarily or possibly answer the
user’s query.
This paper is organized as follows: Section 2 presents a
brief state of the art on aggregation search. Section 3
gives a brief definition of the possibilistic theory. Section
4 is devoted to the description of the model which we
propose. We show, in section 5 an example illustrating
this model. Section 6 gives the evaluation results and
shows the effectiveness of the model. Section 7 concludes
the paper.
Index Terms— XML Information Retrieval, possibilistic
networks, aggregated search, redundancy.
The aim of the aggregated search is to assemble
information from diverse sources to construct responses
including all information relevant to the query. This
comes in contrast with the common search paradigm,
where users are provided with a list of information
sources, which they have to examine in turn to find
relevant content.
It is well known that, in the context of Web search,
users typically access a small number of documents [12].
A study on users Web [24] showed that the percentage of
users who consult fewer documents (Web pages) per
query increases with time. For example, from 1997 to
2001, the users' percentage looking at a document by
query is passed from 28.6% to 50.5%. This percentage
increased further to over 70% after 2001 [25]. It gives to
think that for a list of documents, is mainly confined to
documents in the first, second and sometimes (at most)
third rank. The study reported in [11] showed that on 10
documents displayed, 60% of users have looked less than
5 documents and nearly 30% have read a single
document.
The aggregated search allows to bring solutions to this
problem. Indeed, its aim is to integrate other types of
documents (web documents, images, videos, maps, news,
books) in the results page. Example of search engines that
begin to make aggregation, we find Google Universal
Search, Yahoo! alpha, etc. Users have access to various
document types. This can be beneficial for certain
I. INTRODUCTION
The main problem of content-based XML information
retrieval is how to select the unit of information that
better answers the user’s query [13] [9].
Most of XML Information Retrieval (IR) approaches
[23] [17] [15] [16] [18] consider that the returned units
are a list of disjoint elements (subtrees of XML
documents). We assume that relevant unit is not
necessarily a unique adjoining elements or a document it
could also be any aggregation of elements of that
document. Let us consider a document with the following
structure (document(title)(chappter1(section1)(section2))
chapter2(…)). If the relevant information are located in
the “title” and “section1”, most of XML IR systems will
return the whole document as the relevant unit. In our
case we consider that, the only unit to be returned is an
aggregate (element set) formed by both elements : ”title”
& “section1”. To achieve this objective, we propose a
model enabling to automatically select aggregation of non
redundant elements of the document that better answer
the user’s need formulated through a list of key words.
The model we propose finds its theoretical bases in the
possibilistic networks. The network structure provides a
natural manner to represent the links between, a
document, its elements and its content. As for the
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.181-188
II. STATE OF THE ART
182
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
queries, such as "trip to Finland" can return maps, blogs,
weather, etc.
Another technique that can be used to improve the
search results page is the clustering. However, it is not
enough simply to return the clusters. It is important to
provide users an overview of the contents of the
documents forming a cluster [25]. A common approach to
provide such an overview is a summary of all documents
of the cluster (‘multi documents summary’). Examples of
systems based on this technique, we find NewInEssence
[20], QCS [8], etc.
The issue of elements aggregation from a collection of
XML documents is not addressed in the literature.
Indeed, the proposed approaches that address this issue
are limited to Web documents [6] [1]. However, few
Information retrieval systems begin to aggregate the
results, of a query on XML documents, as summaries.
For example, eXtract [10] is an information retrieval
system that generates results as XML fragments. An
XML fragment is considered a result if it answers four
features: Autonomous (understood by the user), distinct
(different from the other fragments), representative (of
the themes of the query) and succinct. XCLUSTERs [19]
is a model of representation of XML abstracts. It includs
some XML elements and uses a small space to store data.
The objective is to provide significant excerpts for users
to easily evaluate the relevance of query results.
The approach we propose in this paper is located to
junction between the research of the relevant elements
and their regrouping (aggregation) in a same result. Our
approach is based on possibilistic theory [26] [7] [4] and
more particularly on possibilistic networks [2] [3]. These
networks offer a simple and natural model for
representing the hierarchical structure of XML
documents and to handle uncertainty, inherent to
information retrieval. We find this uncertainty in the
concept of document relevance with respect to a query,
the degree of representativeness of a term in a document
or part of documents and the identification of the relevant
part that answers the query. Within this framework, to
identify the relevant part that answers the query, unlike
the approaches suggested in the literature, which select
the sub-tree, likely to be relevant; our approach allows to
identify and to select, in a natural way, an aggregation of
non redundant elements of XML document that may
answer the query.
Besides the points mentioned above, the theoretical
framework, that supports our proposals, namely the
possibilistic networks, clearly differentiate us from the
settings used in previous approaches.
III. THE POSSIBILISTIC THEORY
The possibilistic logic [26] enables to model and
quantify the relevance of a document considering a query
through two measurements: the necessity and the
possibility. The necessarily relevant elements are those
that must appear in top of the list of the selected elements
and must allow system efficiency. The possibly relevant
© 2012 ACADEMY PUBLISHER
elements are those that would eventually answer the user
query.
A. Possibility Distribution
A possibility distribution, denoted by π, corresponds to
a mapping from X to the scale [0, 1] encoding our
knowledge on the real world.
π (x) evaluates to what extent x is the actual value of
some variable to which π is attached. π (x) =1 means
that it is completely possible that x is the real world
(or that x is completely fulfilling), 1> π (x) >0 means
that x is somewhat possible (or fulfilling), and finally
π (x) =0 means that x is certainly not the real world (or is
completely unsatisfactory). An event is said ‘no possible’
does not only mean that the opposite event is possible. It
actually means that it is certain. Two dual measures are
used: the possibility measure П(A) and the necessity
measure N(A).
The possibility of an event A, denoted П(A), is
obtained by П (A) = max x є A π(x) and describes the most
normal situation in which A is true.
The necessity N(A) = min x ∉ A 1 − π(x) = 1 − П(¬A)
of an event A reflects the most normal situation in which
A is false.
B. Product-based Conditioning
In the possibilistic setting, the product-based
conditioning consists of modifying our initial knowledge,
encoded by the possibility distribution π by the arrival of
new fully certain piece of information e. Let us take
Φ = [e] the set of models of e. The initial distribution π is
then replaced by another one π′, such as π′ = π (•/Φ).
Assuming that Φ ≠ Ø and that П (Φ) > 0, the natural
postulates for possibilistic conditioning are:
π(w /p Φ) = π(w)/ П (Φ)
and
0
if w є Φ
otherwise
(1)
Where /p is the product-based conditioning.
C. Possibilistic Networks
A directed possibilistic network over a set of variables
V= {V1, V2,…,Vn} is characterized by:
- A graphical component composed of a Directed
Acyclic Graph (DAG). The DAG structure encodes a set
of independence relations between variables.
- A numerical component consisting in a quantification
of different links in the DAG using the conditional
possibilisties of each node in the context of its parents.
Such conditional distributions should respect the
following normalization constraints for each variable Vi
of the set V:
Let UVi the set of parents of Vi
If UVi = Ø (i.e. Vi is a root), then the a priori possibility
relative to Vi should satisfy:
max ai Π(ai) = 1, ∀ ai ∈ DVi
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
If UVi ≠ Ø (i.e. Vi is not a root), then the conditional
distribution of Vi in the context of its parents should
satisfy:
max ai Π(ai /ParVi) = 1,
∀ ai ∈ DVi
Where ParVi is the set of possible configurations
(aggregations) of parents of Vi
Using the definition of conditioning based on the
product operator. This leads to the following definition of
a product-based possibilistic network:
D. Product-based Possibilistic Network
A product-based possibilistic network over a set of
variables V = {A1, A2,…, AN} is a possibilistic graph
where conditionals are defined using product-based
conditioning (1).
Product-based possibilistic networks are appropriate
for a numerical interpretation of the possibilistic scale.
The possibility distribution of the product-based
possibilistic network, denoted by ΠP, is obtained by the
following product-based chain rule [2]:
ΠP (A1,..., AN) = PRODi=1..N Π (Ai/PARAi)
(2)
Where: ‘PROD’ is the product operator.
IV. THE AGGREGATED SEARCH MODEL
A. Model Architecture
The architecture of the proposed model is illustrated in
Fig. 1. The graph represents the document nodes, index
terms and element nodes (elements of XML document).
The links between the nodes allow representing the
relations of dependences between the various nodes.
T1
E2
T2
En
Ei
Ti
Tj
Tm
Figure 1 . Model Architecture.
Document nodes represent the collection documents.
Each document node Di, represents a binary random
variable taking values in the set dom (Di) = {di, ¬di},
where the value Di = di (resp. ¬di) represents “the
document Di is relevant for a given query (resp. nonrelevant).
© 2012 ACADEMY PUBLISHER
Nodes E1, E2, …, En, represent the elements of
document Di. Each node Ei, represents a binary random
variable taking values in the set dom (Ei) = {ei, ¬ei}. The
value Ei = ei (rep. ¬ei) means that the element ‘Ei’ is
relevant for the query (resp. non-relevant).
Nodes T1, T2, …, Tm are the term nodes. Each term
node Ti represents a binary random variable taking values
in the set dom (Ti) = {ti, ¬ti,} where the value Ti = ti (resp.
¬ti) means that term ‘Ti’ is representative of the parent
node to which it is attached (resp. non-representative of
the parent node to which it is connected). It should be
noticed that a term is connected to the node that includes
it as well as to all its ancestors.
The passage from the document to the representation
in the form of possibilistic network is done in a simple
way. All nodes (elements) represent the level of variables
Ei. The values that will be assigned with the arcs of
dependences between term-element nodes and elementdocument nodes depend on the sense which one gives to
these links.
Each structural variable Ei, Ei ∈ E = {E1, E2, …, En},
depends directly on its parent node which is the root Di in
the possibilistic network of the document. Each variable
of contents Ti, Ti ∈ T = {T1, …, Tm} depends only on its
structural variable (structural element or tag). It should be
also noticed that the representation considers only one
document. In fact, the documents are considered
independent from each other, thus we can consider only
the sub-network, representing the document that is
processed.
We note by T(E) (resp. T(Q)) the set of the index terms
of the elements of the document (resp. of the query).
The arcs are oriented and are of two types:
- Term-element links. These links connect each term
node Ti є T(E) to each node Ei where it appears.
- Elements-document links. These links connect each
element node of the set E to the document that includes it,
in our case Di.
We will discuss interpretations we give these various
links and the way we quantify them, in the following.
Di
E1
183
B. Query Evaluation
As we underlined previously, we model the relevance
according to two dimensions: the necessity and the
possibility of relevance. Our model must be able to infer
propositions of the type:
- “the document di is relevant for the query Q” is
possible to a certain degree or not, quantified by П(Q/di).
- “the document di is relevant for the query Q” is
certain or not, quantified by N(Q/di).
The first type of proposition allows to eliminate the
non-relevant documents, i.e. those that have a weak
possibility. The second proposition focuses the attention
on those that seem very relevant.
For the model presented here, we will adopt the
following assumptions:
184
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Assumption 1: A document has as much possibility to
be relevant than non-relevant for a given user, either
П(di) = П(¬di) = 1,
∀ i.
Assumption 2: The query is composed of a simple list
of key words Q= {t1,t2, … ,tn}. The relative importance
between terms in the query is ignored.
According to the definitions of the possibilistic theory,
the quantities П(Q/di) and N(Q/di) are calculated like
follows :
distinguish the terms possibly representative of the
elements of the document and those necessarily
representative of these elements (terms that are sufficient
to characterize the elements). With this intention, the
possibility of relevance of a term (ti) to represent an
element (ej), noted Π (ti /ej), is calculated like follows:
Π ( t i / e j ) = tf
/max (tf
i j
∀ t k ∈e
kj
)
j
(5)
Π (Q/di ) = max
(Pr ode ( Pr od (Π(ti / θ ej ))) * Pr ode (Π(θ ej / di )) * Π(di ))
e
E
∀θ ∈θ
E j ∈θ
Ti ∈T(E) ^ T(Q)
E j ∈θ
(3)
Where:
- Prod: means product (we used this symbol instead of
∏ not to confuse it with the symbol designating the
possibility).
- ti ∈T(E) ^ T(Q) : represents the terms of the queries
which index the elements of the XML document.
- θ : set of non redundant elements
e
-
θ ej : represents the value of Ej in the aggregation θ e
(example: the value of E1 in the aggregation (e1, e2) is e1).
The selection of the relevant parts (units of
information) is inherent with the model. Indeed, the (3)
calculates the relevance by considering all possible
aggregations (regrouping) of elements. The factor
θe
gives possible values of elements. The aggregation of
elements that will be selected will be the one that
includes obligatorily the terms of the query and presents
the best relevance (maximum relevance) in terms of
necessity and/or possibility.
As it was mentioned in the introduction, our model is
able tri select the best aggregation of elements that are
likely to be relevant to the query. This aggregation is the
aggregation that maximizes the necessity if it exists or the
possibility. It obtained by:
θ * = arg max Π (Q /d i )
∀θ e ∈θ E
(4)
The various degrees Π and N between the nodes of the
network are calculated as follows:
1) Possibility Distribution П(ti/ej)
In Information Retrieval, the terms used to represent
the content of a document, are weighted in order to better
characterize the content of this document. The same
principle is used in XML retrieval. The weights are
generally calculated by using term frequency (tf) within a
document or inverse document frequency (Idf) in the
collection.
In information retrieval, it has been shown [21] [22]
that the performances of the system can be improved if
one represents an element by considering its own content
and the contents of its children nodes. In our model, we
© 2012 ACADEMY PUBLISHER
Where, tfij represents the frequency of the term ‘ti’ in
the element ‘ej’.
A term having a degree of possibility 0 means that the
term is not representative of the element. If the degree of
possibility is strictly higher than 0, then the term is
possibly representative of the element. If it appears with a
maximum degree of possibility, then it is considered as
the best potential candidate for the representation and
thus the restitution of the element.
Let us note that: max (Π (ti / ej)) = 1, ∃ti ∈ ej
In an XML document, a necessarily representative
term of an element is a term that contributes to its
restitution in response to a query. This term is called
discriminative term and it is a term that frequently
appears in few elements of XML document [5]. The
factor commonly used in IR to quantify the
discriminative power of a term is idf (ief in XML IR).
Therefore, a degree of necessary relevance, βij, of the
term ti to represent the element ej, will be defined by:
Ne
N
N(ti →e j) ≥ βij = μ (tfij *iefij) * idf = μ (tfij *log(
) *log(
))
nei +1
ni +1
(6)
Where:
- N and Ne are respectively the number of documents
and elements in the collection.
- ni and nei are respectively the number of documents
and the number of elements containing the term ti.
- μ : a function of normalization. A simple manner to
normalize is to divide by the maximal value of the factor.
- tfij represents the frequency of the term ‘ti’ in the
element ‘ej’.
- iefij represents the inverse frequency of the element
‘ej’ for the term ‘ti’.
- idf represents the inverse frequency of the document.
It should be noticed that (6) has been chosen according
some experiments that were undertaken by Sauvagnat
[22].
This degree of necessary relevance allows limiting the
possibility that the term is compatible with the rejection
of the element by:
Π (ti /¬ej) ≤ 1- βij (this is deduced by definition in the
possibilistic theory)
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
We summarize the possibility distribution on the
Cartesian product {ej, ¬ej} x {ti, ¬ ti} by the following
table:
TABLE I.
ti
POSSIBILITY DISTRIBUTION ON THE SET OF
TERMS T
Π
¬ti
ej
tfij / max(tfkj), (∀tk∈ej)
¬ej
1- βij
1
1
(7)
Where:
- dist(di, ej) is the distance from the element ej to the
root di in accordance with the hierarchical structure of the
document.
- α є ]0..1] is a parameter allowing to quantify the
importance of the distance separating the element nodes
(structural elements of the document) to the root of the
document.
Concerning the necessity to propagate, in an intuitive
manner, one can think that the designer of a document
uses the nodes of small size to emerge important
information. These nodes can thus give precious
indications on the relevance of their ancestors’ nodes. A
title node in a section for example allows locating with
precision the subject of its ancestor node section. It is
thus necessary to propagate the signal calculated on the
level of the node towards the root node. To answer this
intuition, we propose to calculate the necessity of
propagation of relevance of an element ej towards the
root node di, denoted N(e
→ d ) , as follows:
j
equation, the more a term is of small size, the bigger is
the necessity to propagate it.
Therefore, Π (ej/ ¬ di) = lej/dl
We summarize the possibility distribution on the
Cartesian product {di, ¬di} x {ej, ¬ej} by the following
table:
TABLE II.
POSSIBILITY DISTRIBUTION ON THE SET OF
ELEMENTS E
di
Π
2) Possibility Distribution П(ej /di)
The arc document-element (or arc root-element)
indicates the interest to propagate information from an
element towards the document node (root). The nodes,
close to the root (of a tree), carry more information for
the root than those located lower in the tree [22]. Thus it
seems intuitive that more an element is far from the root
more it is less relevant. We model this intuition by the
use in the function of propagation of the parameter
dist(root, e), that represents the distance between the root
and one of its descendant nodes (elements) ‘e’ in the
hierarchical tree of the document, i.e. the number of arcs
separating the two nodes.
The degree of possibility of propagation of relevance
of an element (ei) towards the document node (di) is
defined by Π (ej / di) and is quantified as follows:
Π (ej / di) = α dist(di, ej)-1
185
¬di
ej
α dist(di, ej)-1
lej/dl
¬ej
1
1
V. ILLUSTRATE EXAMPLE
An example of XML document (an extract of a
document) related to a book will be used to illustrate our
talk. The XML document and its possibilistic network are
presented as follows:
<Book>
<Title > Information Retrieval </Title >
<Abstarct> In front of the increasing mass of
information …</Abstract>
….
<Chapter>
<Title chapter> Indexing </title chapter>
<Paragraph> The indexing is the process
intended to represent by the elements of a documentary or
natural language of … </Paragraph>
</Chapter>
</Book>
The possibilistic network associated with XML
document `Book' is as follows:
Book
TitleChapter
Abstrac
Title
Paragrap
Chapter
Information
Retrieval
Indexing
System
XML
i
Figure 2 . Possibilistic network of the XML document ‘Book'
N(e
j
→ d
i
) = 1 −
le
j
dl
(8)
lej is the size of the element node ej and dl the size of a
document (in number of terms). According to this
© 2012 ACADEMY PUBLISHER
For this example, the set of the elements E= {e1=Title,
e2=Abstract, e3=Chapter, e4=Titlechapter, e5=Paragraph}.
The set of the indexing terms of the elements, calculated
while using the content of each element, along with its
186
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
child elements in the document, such as T(E) =
{t1=Retrieval, t2=Information, t3=System, t4=Indexing,
t5=XML}. We consider only some terms not to congest
the example.
The table containing the values of the arcs element
node-term node of the possibilistic network of the
document “Book” is given in Table III. We recall that a
term is connected to the node that includes it as well as to
all the ancestors of this node.
TABLE III.
П (ti/ej)
e1
¬ e1
e2
¬ e2
e3
¬ e3
e4
¬ e4
e5
¬ e5
POSSIBILITY DISTRIBUTION П
t1
t2
1
0
0.5
0.5
0
1
0
1
0
1
1
0
1
0
0
1
0
1
0
1
t3
0
1
1
0
0
1
0
1
0
1
(ti/ej)
t4
0
1
0.25
0.88
0.70
0.10
1
0
0.88
0.05
t5
0
1
0
1
0.5
0.2
0
1
1
0
The table containing the values of arcs root-element
nodes of the possibilistic network of the document
`Book' is given in Table IV (we take α = 0,6 and
dl=100).
TABLE IV.
POSSIBILITY DISTRIBUTION П (ej /di )
П (ej/di)
(di=book)
e1
¬e1
e2
¬e2
e3
¬e3
e4
¬e4
e5
¬e5
1
1
1
1
1
1
0.6
1
0.6
1
П (ej/di)
(di= ¬book)
0.02
1
0.1
1
1
1
0.01
1
1
1
When the query is put, a process of propagation is
started through the network modifying the values of
possibilities a priori. In this model the equation of
propagation used is the (3).
Let’s take a query Q composed of the keywords
“Retrieval”
and
“Information”,
Q={Retrieval,
Information}.
According to the assumption 1, Π (di) = Π(¬ di) = 1,
∀ i.
Given the query Q, the propagation process (3)
considers only the aggregates of set E that include the
query terms t1 = ‘Retrieval’ and t2= ‘Information’. In fact
only the elements e1=’Title’ and e2=’Abstract’ will be
considered. The aggregations that it is thus necessary
considered are: {(e1, e2), (e1, ¬e2), (¬e1, e2), (¬e1, ¬e2)}.
We calculate then:
© 2012 ACADEMY PUBLISHER
For di = book:
a1 = Π( t1/e1) . Π (t2/e1). Π (t2/e2) . Π (e1/book) .
Π (e2/ book) = 1 * 1*1 * 1 * 1= 1
a2 = Π( t1/e1) . Π (t2/e1) . Π (t2/¬e2) . Π (e1/book) .
Π (¬e2/book) = 1* 1*0 * 1 * 1= 0
a3 = Π (t1/¬e1) . Π (t2/¬e1) . Π (t2/e2) . Π (¬e1/book) .
Π (e2/book) = 0 * 0*1 * 1 * 1 = 0
a4 = Π( t1/¬e1). Π( t2/¬e1). Π (t2/¬e2). Π (¬e1/book) .
Π (¬e2/ book) = 0 * 0 * 0 * 1 * 1 = 0
According to (3):
Π(Q/book) = max (a1, a2, a3, a4) = 1 = a1
For ¬ di = ¬ book:
a5 = Π (t1/e1). Π (t2/e1). Π(t2/e2). Π(e1/¬book) .
Π(e2/¬book) = 1 * 1 * 1 * 0.02 * 0.1 = 0.002
a6 = Π (t1/e1). Π (t2/e1). Π (t2/¬e2). Π(e1/¬book) .
Π(¬e2/¬book) = 1 * 1* 0 * 0.02 * 0.1 = 0
a7 = Π (t1/¬e1) . Π (t2/¬e1). Π (t2/e2) . Π(¬e1/¬book) .
Π(e2/¬book) = 0 * 0 * 1 * 1 * 0.1 = 0
a8 = Π(t1/¬e1).Π(t2/¬e1).Π(t2/¬e2) . Π(¬e1/¬book) .
Π(¬e2/¬book) = 0 * 0* 0 * 1 * 1 = 0
According to (3):
Π(Q/¬ book) = max (a5, a6, a7, a8) = 0.002 = a5
The necessity N(Q/book) =1- Π( Q/¬ book) = 1- 0.002
= 0.998
The necessity N(Q/¬ book) =1- ∏( Q/book/) = 1- 1 = 0
The preferred documents are those that have a value
N(Q/di) high among those that have a value Π (Q/di) high
too. If N(Q/di)=0, the restored documents are
(unwarranted of total adequacy) those that have a value
Π(Q/di) high. Therefore, for the query Q = {Retrieval,
Information}, it is the aggregation “a1” (title, abstract)
that will be turned to the user as answer to his query.
VI. EXPERIMENTS AND RESULTS
A. Goals
All studies performed to assess aggregated search were
based on user studies [14].
This user study is designed with two major goals:
- Evaluate the aptitude of an aggregate of XML
elements to answer user queries.
- Identify some of the advantages of the aggregated
search in XML documents.
B. Results
To evaluate our model, a prototype was developed.
Our experiments are conducted on a sample about 2000
XML documents of the INEX'2005 collection, a set of 20
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
queries from the same collection. Every query is assessed
by exactly 3 users.
The following histogram gives the judgments of users
by query regarding the aggregate relevance:
187
Thus, it seems very important to identify other
evaluation criteria to identify all benefits of aggregated
search in XML documents.
REFERENCES
Figure 3 . Distribution of aggregation relevance results
The experimental evaluation shows that aggregated
search has big contribution for XML information
retrieval. Indeed, the aggregate gathers non-redundant
elements (parts of XML document). These elements can
be semantically complementary and in this case the
aggregate allows improving the interpretation of results,
guides the user to the relevant elements of XML
document, faster and also reduces the efforts the user
must provide to locate information searched for.
However, in some cases elements of the aggregate may
be non complementary that means not semantically
related with respect to information need expressed by
user’s query. This sort of aggregation is very useful
because it allows a very fine distinction of the different
thematic expressed in the user's query when his need in
information is generic. It also helps inform the user about
various information of the corpus related to his
information need thus help him, if necessary, to
reformulate his query.
VII. CONCLUSION
This paper presents a new approach for aggregated
search based on possibilitic networks. This model
provides a formal setting to aggregate non-redundant
elements into the same unit. It also directs the user more
quickly toward the relevant elements of XML document.
The user study is constructed around 2 main goals.
First, we analyze the distribution of relevant results
across different elements of XML document. Second, we
identify some of the advantages of aggregated search.
The user study was used to collect supporting data for
these goals. The analysis of the distribution of relevant
results provides interesting information. We notice that
relevant information is sparse across many elements of
XML document.
Our analysis focuses on specific advantages of
aggregated search. It is shown that aggregated search is
useful to identify different interpretations of a query. It
helps find different aspects of the same information need.
© 2012 ACADEMY PUBLISHER
[1] R. Agrawal, S. Gollapudi, A. Halverson, “Diversifying
search results”, ACM Int. Conference on WSDM, 2009.
[2] N. Ben Amor, “Qualitative possibilistic graphical models:
from independence to propagation algorithms”, Thèse
pour l’obtention du titre de Docteur en Gestion, université
de Tunis, 2002.
[3] S. Benferhat, D. Dubois, L. Garcia and H. Prade,
“Possibilistic logic bases and possibilistic graphs”, In Proc.
of the 15th Conference on Uncertainty in Artificial
Intelligence, pp.57-64, 1999.
[4] C. Borgelt, J. Gebhardt and R. Kruse, “Possibilistic
graphical models”, Computational Intelligence in Data
Mining, CISM Courses and Lectures 408, Springer, Wien,
pp.51-68, 2000.
[5] A. Brini, M. Boughanem and D. Dubois, “A model for
information retrieval based on possibilistic networks”,
SPIRE’05, Buenos Aires, LNCS, Springer Verlag, pp. 271282, 2005.
[6] C.L. Clarke, M. Kolla, G.V. Cormack and O. Vechtomova,
“Novelty and diversity in information retrieval evaluation”,
SIGIR’08, pp. 659-666, 2008.
[7] D. Dubois and H. Prade, “Possibility theory”, Plenum,
1988.
[8] D. M. Dunlavy, D. P. O’Leary, J. M. Conroy and J. D.
Schlesinger, “QCS: A system for querying, clustering and
summarizing documents”, Information Processing and
Management, pp. 1588-1605, 2007.
[9] N. Fuhr, M. Lalmas, S. Malik and Z. Szlavik, “Advances
in XML information retrieval: INEX 2004”, Dagstuhl
Castle, Germany, 2004.
[10] Y. Huang, Z. Liu and Y. Chen, “Query biased snippet
generation in XML search”, ACM SIGMOD, pp. 315-326,
2008.
[11] B. J. Jansen and A. Spink, “An Analysis of document
viewing pattern of web search engine user”, Web Mining:
Applications and Techniques, pp. 339-354, 2005.
[12] B. J. Jansen and A. Spink, “How are we searching the
world wide web?: a comparison of nine search engine
transaction
logs”,
Information
Processing
and
Management, pp. 248-263, 2006.
[13] J. Kamps, M. Marx, M. De Rijke and B. Sigurbjörnsson,
“XML retrieval: What to retrieve?”, ACM SIGIR
Conference on Research and Development in Information
Retrieval, pp.409-410, 2003.
[14] A. Kopliku, F. Damak, K. Pinel-Sauvagnat and M.
Boughanem, “A user study to evaluate aggregated search”,
In IEEE/WIC/ACM International Conference on Web
Intelligence, in press.
[15] M. Lalmas, “Dempster-Shafer’s theory of evidence applied
to structured documents: modelling uncertainty”, In
Proceedings of the 20th Annual International ACM SIGIR,
pp.110–118, Philadelphia, PA, USA. ACM, 1997.
[16] M. Lalmas and P. Vannoorenberghe, “Indexation et
recherche de documents XML par les fonctions de
croyance”, CORIA'2004, pp. 143-160, 2004.
[17] P. Ogilvie and J. Callan, “Using language models for flat
text queries in XML retrieval”, In Proceedings of INEX
2003 Workshop, Dagstuhl, Germany, pp.12–18, December
2003.
188
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
[18] B. Piwowarski, G.E. Faure and P. Gallinari, “Bayesian
networks and INEX”, In INEX 2002 Workshop
Proceedings, pp. 149-153, Germany, 2002.
[19] N. Polyzotis and M. N. Garofalakis, “XCluster synopses
for structured XML content”, ICDE, pp. 63, 2006.
[20] D. Radev, J. Otterbacher, A. Winkel and S. BlairGoldensohn, “NewsInEssence: summarizing online news
topics”, Communications of the Association of Computing
Machinery, pp. 95-98, 2005.
[21] T. Rölleke, M. Lalmas, G. Kazai, I. Ruthven and S.
Quicker, “The accessibility dimension for structured
document retrieval”, BCS-IRSG European Conference on
Information Retrieval (ECIR), Glasgow, Mars 2002.
[22] K. Sauvagnat, “Modèle flexible pour la recherche
d’information dans des corpus de documents semistructurés”, Thèse de Doctorat de l’Université Paul
Sabatier, Juillet 2005.
[23] B. Sigurbjornsson, J. Kamps and M. de Rijke, “An
element-based approach to XML retrieval”, INEX 2003
workshop, Dagstuhl, Germany, December 2003.
[24] A. Spink, B. J. Jansen, D. Wolfram and T. Saracevic,
“From e-sex to e-commerce: web search changes”, IEEE
Computer Science, vol. 35, pp. 107-109, 2002.
[25] S. Sushmita and M. Lalmas, “Using digest pages to
increase user result space: preliminary designs”, Special
Interest Group on Information Retrieval 2008 Workshop
on Aggregated Search, 2008.
[26] L. A. Zadeh, “Fuzzy sets as a basis for a theory of
possibility”, In Fuzzy Sets and Systems, 1:3-28, 1978.
Fatma Zohra Bessai-Mechmache
Algiers, Algeria. She obtained her
Engineer Degree in Computer Science
from Houari Boumediene University
(USTHB) in Algeria and her Magister
from the Research Centre of Advanced
Technologies in Algeria.
She has been a team member of the
scientific and research staff of the
Research Centre in Scientific and
Technical Information of Algeria (CERIST) and from 2007 she
is the head of the Databases Team at CERIST. Her research
interests include database security and information retrieval.
She is particularly interested in XML information retrieval,
Aggregated Search and Mobile information retrieval.
Zaia Alimazighi Algiers, Algeria. She
obtained Doctorate in Computer Science
in 1976 at Paris VI University and PHD
in Information System at USTHB
(Algiers’s University) in 1999.
After 1976 and during more than ten
years, she has been project leader in
several industrial projects, in public
companies in Algeria. She is a
researcher at the USTHB since 1988 to nowadays. Today she is
a full prof. at the computer science Department of the Faculty of
Electronic & Computer Science of USTHB and Dean of this
faculty. She is team manager in Computer Engineering
Laboratory at USTHB. The current research interests include
Information Systems, Collaborative Works, Data Warehouses,
Service Web development and Databases Modeling.
© 2012 ACADEMY PUBLISHER
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
189
The Developing Economies’ Banks Branches
Operational Strategy in the Era of E-Banking:
The Case of Jordan
Yazan Khalid Abed-Allah Migdadi
Assistant Professor of Operations Management, Business Administration Department, Yarmouk University
Irbid, Jordan
E-mail: Ymigdadi@yahoo.com
Abstract— The aim of this study is to identify the impact of
adopting e-banking on branches operations strategy during
the period 1999 to 2008, 15 local banks in Jordan were
surveyed by using 3 questionnaires, one was directed to
account operations officer, another to tellers and last one to
branches managers. Annual reports and banks websites
were revised to identify the changes in performance
indicators and e-banking adoption. The study revealed that;
branches are still the main channels of conducting banking
transactions, and e-banking is working in parallel with
branches. This paper is the first paper reports this issue in
developing economies.
Index Terms—E-banking, Branches, Operations Strategy,
Jordan
I.
INTRODUCTION
Banks in developing economies directed toward
adopting e-banking channels during the last decade,
banks in Jordan directed toward adopting a lot of ebanking channels, about 18 out of 23 banks adopted
internet banking, 8 banks adopted telephone banking, 15
adopted mobile banking, and 6 banks adopted ATMs.
This adoption left questions: what is the impact of
adopting e-banking on the branches operations?, and to
what extent these changes affected the performance of
banks?, so answering these questions help the managers
in developing economies to identify the effective role of
branches in the era of e-banking.
Despite the significant adoption of e-banking over the
universe, limited previous studies reported this issue in
developed economies as the study of [40, 25], and no
evidence about this issue in developing economies, so
this study will bridge this gap by reporting this issue in
Jordan. Accordingly the objectives of this study are:
1. Identifying the degree of change in adopting ebanking in Jordan during the period 1999 to
2008.
2. Identifying the degree of change in branches
operations strategy actions and capabilities in
Jordan during the period 1999 to 2008.
3. Identifying the degree of change in banking
performance indicators during the period 1999
to 2008.
4. Identify the impact of significant adopting of ebanking
on
branches
operations
and
performance in Jordan during 1999 to 2008.
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.189-197
II.
OVERVIEW OF BANKING SECTOR IN JORDAN
Jordan is officially known as Hashemite Kingdom of
Jordan, it is a small country located in the Middle East, it
shares boarders with Iraq, Saudi Arabia, and Syria.
Jordanian economy is service oriented, and the banking
sector plays a dynamic role in the Jordanian economy, it
acquiring 56.4% of the total capital invested in the
Amman Stock Exchange [2], further, the banking sector
of Jordan is leader in comparison with its counter parts in
other countries in the Middle East and North Africa
region [15], the number of licensed banks in Jordan is 23
banks (15 local and 8 foreign), the banking sector is
controlled by Central bank of Jordan [3].
During the last decade the banking sector in Jordan
had been reformed, so new banking was launched by mid.
of 2001, electronic banking transaction law was launched
by 2002, and the banks were directed toward compliance
with Basel Accord II by the end of 2008 [24] The banks
in Jordan provide personal and corporate service, the
banks’ branches cover the majority of Kingdom regions.
During the last decade a lot of banks in Jordan have
adopted more e-banking service, more ATM Kiosks have
been lunched, 23 banks have lunched internet banking
during [31], also 13 banks provide telephone banking and
15 banks have lunched mobile banking.
III.
LITERATURE REVIEW
The purpose of literature review is to identify the
branches operational issues that affected by e-banking,
further, identify the banking performance indicators to be
used in evaluating the branches operations strategy.
Limited previous studies tracing of impact of e-banking
on the branches operations in developed economies as the
study of [40, 25], however, no studies envisaged this
issue in developing economies.
Identifying the Operations Actions and Capabilities
Affected by E-banking
A- Branches Location and Accessibility
At the beginning of 1980s, the international trend was
toward decreasing the number of branches as a result of
investing in alternative delivery service channels as
automatic teller machines, and which reduced the
190
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
operating costs, for example the Bank of America closed
one third of their overall branch network, while
increasing the automatic banking machines networks, and
Barclay's of the UK closed 100 of their 3000 full service
branch [8].
Generally speaking the number of bank branches
operating in the UK (excluding North Ireland) has
declined by over 11% from 37,761 in 1983 down to
33,511 in 1993, and building society branches have
similarly declined by over 9% from 6,480 in 1983 down
to 5, 876 in 1993 [17].
Despite the direction of banks during the last few
years toward using remote access distributions channels,
and expansion use of ATM's over the last two decades to
deal with their customer; the number of commercial bank
branches network have been expanded in developed
countries as USA [19, 27, 18], accordingly the banks
during the era of e-banking will concern about the
convenient site location more than past [5].
During 2001 to 2006 brought about the greatest
increase in the bank branches construction in US history,
with nearly 3,000 new branches opening in 2005 [35].
According to Federal Deposit Insurance Corporation, the
number of bank branches increased about 38% to just less
than 70,000 between 1990 and 2004, in 2005 alone the
banking industry had a 3% annual growth rate adding
2,255 net branches offices [16].
The Urban based banks in USA follow their
customers into developing sub-urban, they opened
branches in new retail centers along major highways,
creating advantage over older established competitors
which established their branches in a declining main
street, down town or courthouse square area [35] more
concern about the availability of parking in the front of
the branch, easy to see, low traffic area, and low physical
barriers [13].
However, in Canada the clicks-and-mortar banks'
density in Urban areas is more than sub-urban or rural
area, with closer occurring in urban areas and opening
occurring in sub-urban areas [33], also the direction of
banks to open branches in-store as a part of banks
direction toward more convenience focused delivery
strategy [37].
B- Branches Layout
Management
and
Human
C- Branches Capacity Strategy
The capacity of branches that measured by the
number of employees and teller station could be affected
by adopting E-banking; if the number of branches will be
reduced the same as made by banks in the UK the
capability of branches will be reduced, however, if the
number of branches increased the capacity will be
extended.
Resource
During the era of e-banking the banks operations
manager should start to overhaul branches with a high
percentage of premium customer first, and focus more in
developing the customer experience [21], the branches
will adopt the retail mall layout or the high-touch hightech strategy [10], the focus on this type of layout is on
marketing, or combining better customer service with
expanded opportunity to sell financial products and
improve the relation with customers [13].
External parking area in the front of branch as malls or
restaurants is available, the branch is generally open, the
majority of space will be used for selling and marketing
purposes, further the use of meter or greeting station in
© 2012 ACADEMY PUBLISHER
the front of lobby, this station will assess the customer
needs and direct traffic based on those need, or help
customers specially the old customers to use online
banking [26, 13].
In order to create an environment that encourages
extended interaction between consumers and bank
employees; the bank should move toward more open
teller stations, free standing teller tower that gives
customers the option by standing side by side with the
teller to view information on waist high computer stand,
or seated teller stations; which designed for the customers
who cannot stand at regular teller windows, or for
opening a new account if all closing rooms are occupied
[13], also create area for private conferences a part from
the traffic [6].
The reception area is very comfortable with large
chairs and view into street, current newspapers and
magazines, and a large TV turned to news and financial
programming, further marketing on specially designed
marketing walls and displays, each branch has internet
coffee with fresh coffee or internet kiosks, digital signage
and computer and printer for customers' uses, also some
banks should consider creating a living room in at least
some of their branches [38, 13].
Also the branches will have extensive warm and
natural lighting, natural colors and materials, the lighting
should draw attention to the teller lines, this lighting
system should support productivity and impact how
colors are perceived and how inviting an environment is.
The lobby or branch hall includes some plants pictures,
and the concerns also on availability of central heating
system air conditions venerations, CCTV and Bandit
screens, carpeted floor, the hard service area used in the
front of teller stations, and easy visible signs [6].
D- Transaction Cost and Customers Waiting Time
Using different routes in the branches as ATMs will
reduce the transaction cost [24]; since the number of
employees will be reduced. increasing the process
routings as using e-banking channels in the branches will
reduced the customers waiting time [41], since the
reasonable number of customers will use e-channels so
the number of customers using traditional banking
channel will reduce, so less waiting time in accordance.
Identifying the Performance Indicators of Banking
Operations
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
A- Financial Indicators
The widely adopted financial indicators in the
banking industry are ROA (return on assets) and ROE
(return on equity) [39], and are widely evaluated by
previous studies as the study of [29, 34]. However, the
previous indicators have income from interest and noninterest, so in order to reflect the core banking operations
non-interest income as a percentage of total revenue can
be used [17].
B- Marketing Indicators
The marketing performance can be evaluated to have
better insight about the impact of operations, the widely
adopted indicators by previous studies are 1) market
share [9, 1], 2) growth rate [9], and 3) perceived quality
[36, 29], Firms in recent years have focused more on
tracing retention. Customer retention is a pivotal strategic
issue, and in the 1990s several streams of research
focused on this [11].
IV.
RESEARCH METHODOLOGY
Analytical survey methodology is adopted by this
study as a result of small sample size which is 15 local
banks. The unit of observation is the bank, all banks in
Jordan are the population, however, the local banks (15
banks) were surveyed since the locals are outperform the
foreigners, also, the local banks have larger branches
networks [3]. The time frame of this research is set at ten
years, from the beginning of 1999 to the end of 2008.
During this period the majority of banks started to adopt
certain e-banking elements.
A- The Achievement of First Objective
The adoption of e-banking was traced by revising
banks websites (website archive used for this purpose:
www.archive.org) and annual reports.
Table (1) shows the definition and scales of banking
adoption dimensions. The adoption was measured by
subtracting the adoption points of last period (2008) from
the beginning period 1999 for every bank (Microsoft
Excel was used for this purpose), then, the result was
divided by the adoption during the first period; the result
was the percentage of adopting e-banking.
B- The Achievement of Second Objective:
B-1 Changes in branches accessibility:
The data required to compute the branches
accessibility was collected from bank annual reports and
websites, the number of branches in urban, suburban and
rural areas of every bank was identified from annual
reports. However, the population numbers in urban,
suburban and rural areas were identified from the website
of the department of statistics. After collecting data
accessibility was computed as summarized in table (2),
then, the changes was identified by subtracting the last
© 2012 ACADEMY PUBLISHER
191
period accessibility from the beginning period by using
Microsoft Excel for every bank and every action, then,
the result was divided by the accessibility scores during
the first period; the result was the percentage of change in
branches accessibility.
TABLE (1)
THE MEASUREMENT OF ADOPTING E-BANKING
E-banking Channel
Adoption
Scale
alternatives
Adoption of ATMs
ATM
Urban Accessibility:
Number
of
Accessibility.
ATM Suburban ATM/10,000
People.
Accessibility.
ATM
Rural Models: each
models
1
Accessibility.
Number of ATM point.
Models.
Adoption
of
Internet Number
of 1 point each
service
Banking.
Transactional
Services.
Adoption
Banking
Adoption
Banking
of
of
General
Adoption
Mobile
Telephone
E-banking
SMS Push
SMS Pull
Mobile Internet
Banking.
IVR (Interactive
Voice Response)
Call Centre.
Contact Centre.
ATM
Internet Banking
SMS Push
SMS Pull
Mobile internet
banking
IVR
Call Centre
Contact Centre
1 point each
service
1 point each
service
1 point each
service
TABLE (2)
BRANCHES ACCESSIBILITY ALTERNATIVES AND SCALES
Branches
Accessibility
References
Scale
Branches
urban
accessibility
Branches
suburban
accessibility
Branches
rural
accessibility
[32, 14]
Number of branches
/10,000 people
B-2 Changes in branches layout design and
number of tellers stations
Primary data was collected from branch managers’ inbranch, since they have direct contact with branch layout
daily, and so are capable to provide historical data about
layout design. Branch managers in different geographic
regions were surveyed, since branch-layout design may
be change from region to region; therefore multiple
respondents from different geographic regions increased
the reliability of data.
Data was collected using a questionnaire developed by
the researcher, and the items were developed according to
literature, the number of items in this questionnaire was
192
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
57 (see table (3)). Data was collected from convenient
sample of branches managers of 15 banks, who have been
appointed in this position since 1999, the total number of
participated managers in this survey was 236 (response
rate 68%). After collected the facts the data was coded by
using competency score presented in table (3), for each
bank, and the data was aggregated for every bank by
computing the average responses.
The changes in each sub-dimension for every bank
were identified by subtracting the competency score of the
last period from the beginning period (Microsoft Excel
was used for this purpose) then, then, the result was
divided by the layout scores during the first period; the
result was the percentage of change in layout design.
B-3 Changes in Tellers Flexibility, Account
Customers Waiting Time:
Tellers were surveyed to identify whether there is a
change in their role (directing toward selling and
promoting products), and customers waiting time. Tellers
were source of these facts since they are in direct
interaction with operations. Data was collected from
convenient sample of tellers of 15 banks, who have been
appointed in this position since 1999, the total number of
participated tellers in this survey was 241 (response rate
70%). Questionnaire was developed by researcher, after
collecting data the responses were coded as presented in
table (4). The customers waiting time was measured by
minutes. However, tellers flexibility was measured as
summarized in table (4).
TABLE (4)
TELLERS FLEXIBILITY: QUESTIONNAIRE’S ITEMS AND
SCALES
Flexibility
Number
Items
Scale
Dimensions
of items
References
Promotion
1
[22, 23,30, Yes: 1
28]
Making
2
solve problems:
Decisions
1, making interest
decision: 2
Cross Trained
2
Do back office: 1,
do front office: 1
B-4 Transaction Cost and number of tellers
Account transaction cost and number of tellers were
identified from account operations unit in the headquarter
by surveying account executive officer who have been
appointed in this position since 1999 (one officer in each
bank) so 15 executives were surveyed, since this unit has
records about this fact, further, the branches managers
and tellers do not have idea about such facts. Data was
collected by using a questionnaire. The scale of
transaction cost was currency unit (JD) and for number
teller (number). The changes were identified in all of
these dimensions by subtracting the score of the last
period from the beginning period by using Microsoft
Excel for every bank then, then, the result was divided by
the score during the first period; the result was the
percentage of change in transaction cost and number of
tellers.
© 2012 ACADEMY PUBLISHER
C. The Achievement of Third Objective
The source of all of these indicators was banks’
annual reports except customer satisfaction and retention
which were collected from headquarter by surveying
account operations officer. Then the change over the
study period were identified by subtracting the results of
last first period from last period, then the relative
performance of each bank according to each performance
indicators was identified by dividing the banks’ change in
performance indicator over the best performed bank.
D. The Achievement of Fourth Objective
K-means cluster analysis was used to cluster banks
according to degree of adopting e-banking, all clustering
trial were generated until reach the maximum number of
clustering trial, since the number of clusters were not
known. Then, the clusters that adopted significant ebanking practices were identified by using Kruskal Wallis
H test, since data is not normally distrusted and sample
size is small.
The significant branches operations strategy actions,
capabilities and performance indicators of each clustering
trial adopted significant e-banking practices were
identified by using Kruskal Wallis H test. The level of
significance was 0.05.
The significant e-banking and branches operations’
practices were represented by using actual scores (the
scores were represented in table (1,2,3,4) in previous
section), these were identified by subtracting the
indicators of the beginning period from the last period.
Further the percentage of change was identified the same.
V. DATA ANALYSIS AND FINDINGS
Table (5) shows significant e-banking practices across
clustering trials, it can be seen that; the significant ebanking practices were identified in 3rd and 4th clustering
trial only. The significant practices across clusters of 3
clusters trial were; ATM suburban accessibility only,
however, no significant practices related to internet
banking, mobile banking and telephone banking.
The significant practices across clusters trial of 4
clusters trial were; ATM suburban and degree of adopting
mobile banking, no significant practices of
internet banking, and telephone banking. The average
ATM accessibility across 3 clusters trial was 1.28
ATMs/10,000 people, and the accessibility across 4
cluster trial was 1.19 ATMs/10,000 people. The
maximum accessibility of 3 cluster trial was reported by
cluster 2 (3.34 ATMs/10,000 people), and the maximum
accessibility of 4 clusters trial was reported by cluster 3
(3.53 ATMs/10,000 people). The significant mobile
banking practices of 4 cluster trial was adopting push
SMS banking, the best cluster was 3rd cluster which
adopted pull SMS banking.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
193
TABLE (3)
BRANCHES LAYOUT DESIGN: QUESTIONNAIRE’S ITEMS, SCALES AND COMPETENCY SCORES
Dimensions of Main
Sub-dimension
Number of items
Item References
Competency Score
Actions
(Items in
Questionnaire)
Visual
-----3
[4, 7, 13, 15]
Window: 1 teller
stations: 3, lighting: 5
Isolation
2
[4, 7, 13, 15]
Yes: 1 No: 0
Air conditioning
5
Central heating
2
Seats and disks
3
Parking
Colours
1
3
Floor
2
Pictures and plants
CCTV
1
2
Alarms
3
Recycling
3
Walls
2
Community
2
Signs and labels
5
Promotional leaflets
and facilities
3
Badges and uniforms
Customer service
2
2
Departmentalisation
2
Internet banking
2
ATM
2
Type of teller stations
3
Number
1
Convenience
Aesthetic
Safety and security
[15]
[13, 7]
Social responsibility
Information factor
E-banking channels
Teller stations
© 2012 ACADEMY PUBLISHER
[4, 6, 13, 15]
[4, 13]
[4, 13]
Ventilation: 1,
entrance air
conditioner: 3, hall air
conditioner: 5, teller
station air
conditioner: 7, credit
offices air
conditioner: 9
Central heating: 1
Heating thermostat: 2
Seats in the front of
teller stations: 1, VIP
halls: 5, disks: 7
Yes: 1 No: 0
Warm colours: 3,
cool colours: 2,
subdued colours: 1
Carpeted floor: 1,
hard surface: 2
Yes: 1, No: 0
CCTV in the branch:
1, outside: 2
Entrance: 1, front of
teller stations: 3,
offices: 5, connected:
7
Solar panel: 5,
collecting water: 3,
recycled rubber floor:
1
Changed colours: 1,
classed walls: 2
Children's playing
area: 1, community
meeting halls: 2
Dept. title signs: 5,
direction signs: 3,
teller station title and
number signs: 1
Digital signs: 2,
traditional: 1
Leaflets at the
entrance: 1, leaflets in
the hall: 3, TV
screens: 5
Uniform: 2, badges: 1
Electronic kiosks: 2,
staff: 1
Departmentalisation:
1, customer service
unit: 2
Wireless laptop area:
2, Internet café: 1
Wall ATMs: 1, inside
ATM: 2
Stand-up: 1, seat: 3,
tower: 5
Number
Max. Score
9
1
25
3
13
1
3
3
1
3
16
15
9+2
5
3
3
3
3
3
9
Relative
number > Average +
standard deviation
194
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Cluster
number
Across all
clusters
Significant
banking
practices
e-
ATM Suburban
Accessibility
Cluster 1
ATM Suburban
Accessibility
Cluster 2
ATM Suburban
Accessibility
Cluster 3
ATM Suburban
Accessibility
Cluster 4
-------
TABLE (5)
SIGNIFICNAT E-BANKING PRACTICES ACROSS CLUSTERING TRIALS
3 clusters trial
4 clusters trial
Percentage of
Percentage of Actual
Significant e- Kruskal
Kruskal
Change
banking
Wallis
H change
Wallis
H change
practices
test
test
Chi.
Chi.
Sig.
Sig.
5.21%
ATM Suburban 10.406
5.165
6.39%
1.28
0.015
Accessibility
0.023
ATMs/
10,000
Mobile Banking
people
Adoption
1.44%
8.775
0.023
1.78%
ATM Suburban --------0.21%
0.35
Accessibility
ATMs/
10,000
Mobile Banking
people
Adoption
0.00
15.62%
ATM Suburban --------3.34%
3.34
Accessibility
ATMs/
10,000
Mobile Banking
people
Adoption
0.00
-----
-----
15.62%
-----
0.84
ATMs/
10,000
people
-----
ATM Suburban
Accessibility
Mobile Banking
Adoption
ATM Suburban
Accessibility
Mobile Banking
Adoption
Table (6) shows significant branches operations
actions across clustering trials, it can be seen that; the
significant actions identified in 3rd clusters trial were;
trained front office employees to do back-office job and
reduce customers waiting time, however, the significant
actions identified in 4th clusters trial was trained frontoffice employees to do back-office jobs, so no significant
actions related to branches accessibility, branches layout
quality, and transaction cost.
Cluster
number
Across
clusters
Cluster 1
all
-----
-----
1.19
ATMs/
10,000
people
0.68: Push
SMS
1.78
ATMs/
10,000
People
0
0.84
ATMs/
10,000
people
0
3.53%
3.21ATMs/
10,000
people
1.75%
1.75: Pull
SMS
-0.47
ATMs/
10,000
people
-0.08%
1.13%
1:
push
SMS
The average percentage of employees trained to do
back-office job across 3 clusters trial was 56%, and the
percentage across 4 cluster trial was 66% of employees.
The maximum percentage of trained employees of 3
cluster trial was reported by all clusters except fourth
cluster (100% of employees). The average reduction of
customers waiting time of 3 cluster trial was reducing
customers waiting time by 7.9 minutes; the best cluster
was 3rd cluster, the customers waiting time was reduced
by 9.5 minutes.
TABLE (6)
SIGNIFICNAT BRANCHES OPERATIONS ACTIONS ACROSS CLUSTERING TIRALS
3 clusters trial
4 clusters trial
Percentage
Percentage
Actual
Significant e- Kruskal
Kruskal
Significant
Wallis H of change
Change
banking
Wallis H of change
branches
test
practices
test
operations actions
Chi.
Chi.
Sig.
Sig.
70%
56%
of Trained front 11.345
61%
6.051
Trained
front
0.010
office
employees
0.014
office employees
employees to
to do back-office
do back-office
job.
job.
Reduce
account
-7.9 minutes
-63.4%
3.710
customers waiting
0.05
time
Trained
front
-----16.7%
-16.7%
of Trained front
----100%
© 2012 ACADEMY PUBLISHER
Actual
Change
Actual
Change
66%
of
employees
100%
of
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
employees
office employees
to do back-office
job.
Cluster 2
Reduce
account
customers waiting
time
Trained
front
office employees
to do back-office
job.
Cluster 3
Reduce
account
customers waiting
time
Trained
front
office employees
to do back-office
job.
Cluster 4
Reduce
account
customers waiting
time
-------
-----
-----
-----
-41.5%
-7.03
minutes
100%
88%
of
employees
-73%
-7.72
minutes
100% of
employees
100%
of
employees
-76%
-9.5 minutes
-----
-----
Table (7) shows the differences in performance
indicators across clustering trial, it can be seen that; no
significant differences in performance indicators across
195
employees
office
employees to
do back-office
job.
Trained front
office
employees to
do back-office
job.
-----
100%
100% of
employees
Trained front
office
employees to
do back-office
job.
-----
83%
100% of
employees
Trained front
office
employees to
do back-office
job.
-----
-18.8%
18.8-% of
employees
clusters in 3 and 4 clusters trials, so, the significant
changes in e-banking and branches operations did not
lead to significant changed in performance.
TABLE (7)
KRUSCAL WALLIS H TEST FOR DIFFERENCES IN PERORMANCE INDICATORS ACROSS CLUSTERING TRIALS
Performance Indicators
3 clusters trial
4 clusters trial
Chi-square
Sig.
Chi-square
Sig.
Return on Equity
1.900
0.593
0.000
1.000
Return on Assets
2.657
0.448
0.320
0.572
Operating revenue/total revenue
3.244
0.356
2.337
0.126
Deposits Market share
1.250
0.741
0.080
0.777
Customers Satisfaction
4.447
0.217
0.082
0.774
Customers Retention
1.998
0.573
0.139
0.710
VI. DISCUSSION
ATM suburban accessibility was significant ebanking practices since ATM has deployed in Jordan
since mid. 1990s, but the majority of banks have ATMs
in urban areas, so it is more attractive to launch ATMs in
suburban. However, significant mobile banking practice
is reasonable since the majority of banks adopted Push
SMS so adopting Pull SMS is the significant.
No significant operations related to branches location,
branches layout, tellers role in promoting or selling
products, number of teller stations and number of tellers
which indicates that the branches are still working in
parallel with e-banking channels, so the branches are still
a channel of conducting transaction rather than a place of
promoting and selling products, accordingly, no
significant impact on transaction cost.
The significant training of employees to do backoffice tasks is reasonable since the direction of customers
toward using e-banking channels reduce the work load on
front-office employees so they may be have more time to
© 2012 ACADEMY PUBLISHER
do back office job, further, the reduction of customers
waiting time is as a result of using customers e-banking
channels which reduce number of served customers by
tellers, and reduce customers waiting time in accordance.
No significant impact of adopting e-banking on
performance, since e-banking practices are still recent in
Jordan so the impact could be significant in future.
VII. CONCLUSION
This study investigated the impact of adopting ebanking on traditional banking operations in Jordan. Ebanking practices traced were; degree of adopting ATMs,
internet banking, telephone banking, mobile banking and
e-banking channels. A scale was developed in this study
to measure the degree of adoption. Traditional banking
operations actions and capabilities were traced in this
study; branches accessibility, branches layout quality,
tellers flexibility, transaction cost, customers waiting
time, and branches capacity.
The practices of 15 local banks in Jordan were
reported by using questionnaires directed to tellers,
196
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
account operations executives, branches managers,
further, annual reports were revised. K-means Cluster
analysis was used to identify the clusters of banks
adopted significant e-banking practices; the significant
practices were adopted by three and four cluster trials,
Kruskal Wallis H test used to identify significant
differences across clusters.
The significant e-banking practices adopted by banks
in Jordan were; ATM suburban accessibility and adopt
mobile banking, however, the significant traditional
banking operations actions and capabilities were; reduce
customers waiting time and trained tellers to do backoffice job. No significant change in branches
accessibility, branches layout design, transaction cost,
and branches capability were reported in Jordan which
indicates that branches are still the main channels of
conducting banking transactions, and e-banking is
working in parallel with branches.
VIII. APPLICATIONS AND FUTURE RESEARCHES
The result of this is study is important for decisions
makers and academics alike. The decision makers of
banks in Jordan are having now facts about the
effectiveness of actions they made in era of e-banking, so
they can focus on the effective actions made and those
not adopted some of these actions can plan now to realize
these actions in future. Further, the proposed investors
who are concerning about banking in Jordan can now
make more rational actions.
Academics now have some facts about the impact of
e-banking on traditional banking operations strategy in
developing countries, so these facts could be used to
develop some propositions and hypothesis, further,
academics now have some facts about banking process to
be taught to students at universities, which help students
in developing countries to have better insight about the
actions they should be made in future.
Despite Jordan is one of the developing countries, the
results of this study is more applicable for banks in
Jordan, so it is recommended to conduct further studies in
other developing countries to find whether the same
significant actions and capabilities are adopted by banks
and trace the impact of actions on capabilities. The future
studies could be more beneficial if it is cross countries or
regions studies, since a comparison can be made and
better conclusions about the impact of country or region
context could be traced.
REFERENCES
[1] Ahmed, N. and Montagno, R. (1995), "Operations strategy
organizational performance: an empirical study",
International Journal of Operations and Production
Management, Vol. 16 No.5, pp. 41-53.
[2] Amman Stock Exchange, (2006), Capital market profile
2006 [online], Available from: http://www.exchange.jo,
[accessed at 20 Feb. 2008].
[3] Association of Banks in Jordan (2007), Annual Report 29th
[online].
Available
from:
http://www.abj.org.jo/Portals/0/AnnualReport/AnnualRep
ort2007.pdf, [Accessed at 20th July 2008].
© 2012 ACADEMY PUBLISHER
[4] Baker, J. Berry, L.L. and Parasuraman, A. (1988), "The
marketing impact of branch facility design", Journal of
Retail Banking, Vol.10 No.2, pp. 33-42.
[5] Beery, A. (2002), "Site selection insight", ABA Banking,
Vol. 34 No. 4, pp. 28-32.
[6] Belski, L. (2004), "Seven design problems and a few good
ideas on improving interiors", ABA Banking Journal,
Vol.96 No.8, pp. 28-32.
[7] Bielski, L. and Street, B. (2007), "Evoking a sense of place",
ABA Banking Journal, Vol. 99(8), pp. 25-30.
[8] Channon, D.F. (1986) Bank strategic management and
marketing, Wiley and Sons, United Kingdome.
[9] Cleveland, G. (1989), "A theory of production competency",
Decision Science, Vol.20 No.4, pp. 655-668.
[10] Davidson, S. (2002), "Bank branch design helps improve
the bottom line", Community Banker, Vol. 11 No. 3, pp.
38-39.
[11] Erikssan, K. and Vaghutt, A.L. (2000), "Customer
retention, purchasing behavior and relationship substance
in professional services", Industrial Marketing
Management, Vol. 29, pp. 363-372.
[12] Creane, S. Goyal, R. Mobarak, A.M. and Sab, R. (2004),
Financial sector development in the Middle East and
North Africa, IMF working paper [online], Monetary
Fund, Middle East and Central Asia Department,
Washington,
DC
USA.
Available
from:
http://www.imf.org/external/Pubs/Pubs/and-inf.htm
[accessed at 2 August 2008].
[13] Feig, N. (2005), "Extreme (and not so extreme) makeovers
in modern branch design", Community Banker, Vol. 14
No. 5, pp. 32-40.
[14] Frohlich, M. and Dixon, J. (2001), "A taxonomy of
manufacturing strategies revisited", Journal of Operations
Management, Vol. 19, pp. 541-558.
[15] Greenland, S. and McGoldrick, P. (2005), "Evaluating the
design of retail financial service environments", The
International Journal of Bank Marketing, Vol.23 No.2/3,
pp. 132-152.
[16] Grover, C. and Ferris, J. (2007), "Right location wrong
building", Community Banker, Vol. 16 No. 10, pp. 48-52.
[17] Hallowell, R. (1996), "The relationship of customer
satisfaction, customer loyalty, and profitability: an
empirical study", International Journal of Service
Industry Management, Vol.7 No.4, pp.27
[18] Hirtle, B. (2007), "The impact of network size on bank
branch performance", Journal of Banking and Finance,
Vol. 31, pp. 3782-3805.
[19] Hirtle, B. and Melti, C. (2004), "The evaluation of US bank
branch network: growth, consolidation, and strategy",
Current Issues in Economic and Finance, Vol. 10 No. 8.
[20] Howcroft, B. and Beckett, A. (1996), "Branch network and
the retailing of high credence products", The International
Journal of Bank Marketing, Vol.14 No.4, pp.4-6.
[21] Hoeck, B. (2006), Customer centric banking 2007:
customer relationship redefined-how retail banks can
learn from other industries [online], White paper: GFT
technology
AG:
Germany,
Available
from:
http//www.GFT.com, [accessed at May 3 2008].
[22] Hunter, L. and Hittle, L.M. (1999), What makes a highperformance workplace?: evidence from retail bank
branches [online], Financial Institution Centre: The
Wharton School: University of Pennsylvania: USA.
Available from: http://www.Watonschool.com [accessed
at 3rd Dec. 2008].
[23] Hunter, L. (1995), How will competition change human
resource management in retail banking?: a strategic
perspective [online]. Financial Institution centre: The
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Wharton School: University of Pennsylvania: USA.
Available from: : http://www.Watonschool.com [accessed
at 3rd Dec. 2008].
[24] Khamis, M. (2003), Financial Sector reforms and issues in
Jordan [online]. The Euro-Med Regional Economic
Dialoge: Rome, October 20. Available form:
http://eceuropa.eu/external_relations/euromed/en/red/kha
mis.pdf, [assessed at March 10 2008].
[25] Lymperopoulos, C. Chaniotakis, I. (2004), “Branch
employees' perceptions towards implications of e-banking
in Greece”, International Journal of Retail & Distribution
Management, Vol. 32 No.6/7, pp.302-311.
[26] Lombardo, T. (2003), "Facility design plays key role in
supporting brand identity", Bank News, Vol. 103 No. 8,
pp. 20-21.
[27] Lopez, R.H. and Shah, J. (2004), Branch Development in
an era of commercial banking consolidation (1994-2006)
[online], Pace University: Lubin School of Business,
USA,
Available
from:
http//digitalcommons.pace.edu/lubinfaculty_working
papers/59, [accessed at February 24 2008].
[28] McKendrick, J. (2002) “Back to the branch: customer may
demand access at every e-channel but they still want a
smiling face”, Bank Technology News, Vol.15 No.5, p.1.
[29] Menor, L. Roth, V. and Mason, C. (2001), "Agility retail
banking: a numerical taxonomy of strategic service
groups", Manufacturing and Service Operations
Management, Vol. 13 No. 4, pp. 273-292.
[30] Metter, R. and Vargas, V. (2000), "A typology of decoupling strategies in mixed services", Journal of
Operations Management, Vol. 18, pp. 663-682.
[31] Migdadi, Y.K.A. (2008), "the quality of internet banking
service encounter in Jordan", Journal of Internet Banking
and Commerce [online], December, 13(3), Available
form:
http://http://www.arraydev.com/commerce/jibc/,
[Accessed at 13th March 2008].
[32] Miller, J. and Roth, A. (1994), "A taxonomy of
manufacturing strategies", Management Science, Vol.
40 No. 3, pp. 285-304.
[33] Mok, M.A. (2002), From bricks-and-mortar to Bricks-andclicks: the transformation of a bank branch network
within greater Toronto area, Master thesis, Wielfrid
Laurier University: Canada.
© 2012 ACADEMY PUBLISHER
197
[34] Power, T. and Hahn, W. (2004), "Critical competitive
methods, generic strategies and firm performance", The
International Journal of Bank Marketing, Vol. 22 No. 1,
pp. 43-64.
[35] Reider, S. (2006), "The branching boom", Kentucky Banker
Magazine, Jun, pp. 18-19.
[36] Roth, A. and Jackson, W. (1995), "Strategic determinants
of service quality and performance: evidence from the
banking industry", Management Science, Vol.41 No.11,
pp. 1720-1733.
[37] Stroup, C. (1998), "Beyond in-store banking at star bank",
Journal of Retail Banking Services, Vol. 20 No. 4, pp. 1925.
[38] Swann, J. (2006), "Let the branch be unbroken",
Community Banker, Vol. 15 No. 10, pp. 58-62.
[39] Uzelac, N. and Sudarevic, T. (2006), "Strengthening the
link between marketing strategy and financial
performance", Journal of Financial Service Marketing,
Vol.11 No.2, pp. 142-156.
[40] Yakhlef, A. (2001), “Does the Internet compete with or
complement
bricks-and-mortar
bank
branches?”,
International Journal of Retail & Distribution
Management, Vol. 29 No. 6/7, pp.274-28.
[41] Klassen, K. and Rohleder, T. (2001), "Combining
operations and marketing to manage capacity and demand
in services", The Service Industry Journal, Vol. 21 No.2,
pp. 1-30.
Dr. Yazan Khalid Abed-Allah Migdadi: Assistant
professor of operations management, Yarmouk University,
Jordan. He awarded Ph.D. in operations strategy from Bradford
University, UK, also awarded BA and MBA from Yarmouk
University, Jordan. He worked as teaching assistance and
researcher of operations and information management at
Bradford University, School of Management, UK. And a
lecturer of Management at Yarmouk University. His research
interests concentrated on operations strategy; reporting best
practices in developing economies, reporting taxonomies of
operations strategies in developing economies and developing
typologies of operations strategy.
198
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
Evolving Polynomials of the Inputs for Decision
Tree Building
Chris J. Hinde and Anoud I. Bani-Hani
Loughborough University/Computer Science, Loughborough, UK
Email: {C.J.Hinde, A.I.Bani-Hani}@lboro.ac.uk
Thomas.W. Jackson
Loughborough University/Information Science, Loughborough, UK
Email: T.W.Jackson@lboro.ac.uk
Yen P. Cheung
Monash University/Clayton School of IT, Wellington, Australia
Email: Yen.Cheung@monash.edu.au
Abstract— The aim of this research is to extend the discrimination of a decision tree builder by adding polynomials of
the base inputs to the inputs. The polynomials used to
extend the inputs are evolved using the quality of the
decision trees resulting from the extended inputs as a fitness
function. Our approach generates a decision tree using the
base inputs and compares it with a decision tree built using
the extended input space. Results show substantial
improvements. Rough set reducts are also employed and
show no reduction in discrimination through the
transformed space.
Index Terms—Decision tree building, Polynomials
I. INTRODUCTION
This paper addresses the well-known problem of data
mining where given a set of data; the expected output is a
set of rules. Decision trees using the ID3 approach [1],
[2] are popular and in most cases successful in generating
rules correctly. Extensions to ID3 such as C4.5 and
CART are developed to cope with uncertain data. Fu et al
[3] used C4.5 followed by a Genetic Algorithm (GA) to
evolve better quality trees; in Fu’s work C4.5 was used to
seed a GA, which were then used as a basis for evolving
better trees then using Genetic Programming (GP)
techniques to cross over the trees. Many rule discovery
techniques combining ID3 with other intelligent
techniques such as genetic algorithms and genetic
programming have also been suggested [4], [5], [6].
Generally, when using ID3 with genetic algorithms,
individuals which are usually fixed length strings are
used to represent decision trees and the algorithm evolves
to find the optimal tree. When Genetic Programming is
used to generate decision trees, individuals are variable
length trees, which represent the decision tree. Variations
in these approaches can be found in the gene encoding.
One rule per individual as done in Greene [7], Freitas et
al [8], [9] is a simple approach but the fitness of a single
rule is not necessarily the best indicator of the quality of
the discovered rule set. Encoding several rules in an
individual requires longer and more complex operators
© 2012 ACADEMY PUBLISHER
doi:10.4304/jetwi.4.2.198-203
[10], [11]. In genetic programming, a program can be
represented by a tree with rule conditions and/or attribute
values in the leaf nodes and functions in the internal
nodes. Here the tree can grow dynamically and pruning
of the tree is necessary [12]. Papagelis & Kelles [13] used
a gene to represent a decision tree and the GA then
evolves to find the optimal tree, similar to Fu et al [3]. To
further improve the quality of the trees, Eggermont et al
[14] applied several fitness measures and ranked them
according to their importance in to tackle uncertain data.
Previous work has taken the input space as a given and
used evolution to produce the trees. In this work, as we
shall see, the trees are generated using a variant of C4.5
and the input space is evolved rather than the trees, in
direct contrast to other workers.
A vast majority of the approaches use decision trees as
a basis for the search in conjunction with either a GA or
GP to further improve the quality of the trees. Our
approach described in this paper addresses continuous
data and adds polynomials of the input values to extend
the input set. A GA is used to search the space for these
polynomials based on the quality of the tree discovered
using a version of C4.5.
II. ITERATIVE DISCRIMINATION
ID3, C4.5 and their derivatives proceed by selecting an
attribute that results in an information gain with respect to
the dependent variable. A simple data set with 2
continuous attributes that are linearly separable is shown
in Fig. 1.
Applying C4.5 to the data set gives the result shown in
Figure 2, which was first documented in [15]. If no errors
are required over a large training set then the complexity
of the decision tree grows with the size of the training set.
This is unsatisfactory.
Anticipating the results of the proposed system a higher
level discriminant of x−y in addition to the two basic
variables x and y would give the result shown in Fig. 3.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
199
were deemed outside the concept. Fig. 4 illustrates the
data set although, as above, does not show all the points.
Applying C4.5 to the data set represented in Fig. 4
gives the decision tree shown in Fig. 5. This decision tree
is smaller than the decision tree derived from the linearly
separable data although the function used to produce the
data is much more complex, and the predictions from the
tree show fewer errors. The decision tree is difficult to
interpret.
Fig. 1. A granularised version of the linearly separable set of data based
on a 2 dimensional data set.
x <= -0.25 :
|
y > -0.75 : in (36.0)
:
:
:
x > -0.25 :
|
y <= 0.75 : out (40.0)
|
y>0.75:
:
:
:
Evaluation on training data (128 items):
Before Pruning
------------------Size Errors
27 1( 0.8%)
After Pruning
-----------------------Size Errors
Estimate
27 1( 0.8%) (13.3%)
Fig. 2. The decision tree produced by C4.5 from the linearly separable
data shown in Figure 1. The size of 27 indicates why this tree is not
replicated here.
x-y <= -0.5 : out (64.0)
x-y > -0.5 : in (64.0)
Evaluation on training data (128 items):
Before Pruning After Pruning
------------------- -----------------------
Fig. 4. A granularised version of the torus illustrating a
quadratic form.
x <= -3.25 : out (16.0)
x > -3.25 :
|
x > 2.75 : out (16.0)
|
x < 2.75 : out (16.0)
|
|
y <= -3.25 : out (12.0)
|
|
y > -3.25 :
|
|
|
y <= 2.75 : in (72.0/24.0)
|
|
|
y > 2.75 : out (12.0)
3
Evaluation on training data (128 items):
Before Pruning After Pruning
-------------- -----------------------Size Errors Size Errors Estimate
9
24(18.8%)9 24(18.8%)(25.5%)
Fig. 5. The decision tree produced by C4.5 from the toroidal data.
Size Errors
3
0( 0.0%)
Size Errors Estimate
3 0( 0.0%) ( 2.1%)
Fig. 3. The decision tree produced by C4.5 from the linearly separable
data using the discriminant value x − y.
III. MORE COMPLEX DISCRIMINANTS
So far we have made no more progress than Konstam
[16] who used a GA to find linear discriminants. He
makes the comment that the technique can be applied to
quadratic discriminants. However he makes no statements
about focusing the search. A set of data was prepared
using the same data points as above to explore higher
order and higher dimensional discriminants. The data
prepared used a torus such that points inside the torus
were in the concept and points outside the torus,
including those that are within the inner part of the torus,
© 2012 ACADEMY PUBLISHER
Taking the toroidal data set, Fig. 4, and adding another
attribute computed from the sum of squares of x and y
gives better discrimination and a more interpretable tree
shown in Fig. 6. Notice that the decision tree is much
smaller with 5 decision points compared to 9, and has no
errors compared with 18.8% in the original tree, Fig. 5.
IV. NON PROJECTABLE DATA SETS
Thus far we have seen data sets that can be projected
onto 1 dimension and which result in large trees but are
nonetheless useful predictors. Section III shows that these
trees can be reduced in size considerably by adding
higher dimensional combined functions of the original
data elements.
200
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
r2 > 8.125 : out (72.0)
r2 <= 8.125 :
|
r2 <= 0.625 : out (8.0)
|
r2 > 0.625 : in (48.0)
Evaluation on training data (128 items):
Before Pruning After Pruning
------------------- -----------------------Size Errors
Size Errors Estimate
5
0( 0.0%)
5 0( 0.0%) ( 3.1%)
Fig. 6. The decision tree produced by C4.5 from the augmented
toroidal data. r2 is the sum of the squares of x and y.
This test shows a data set that cannot be discriminated
by C4.5, however a decision tree does exist. It is shown if
Fig. 10. This clear 2 dimensional data set results in the
following tree from C4.5, Figure 12.
V. GENETIC ALGORITHM
The genetic algorithm attached to the front of c4.5 has
a few special features. It follows most of the guidelines in
[17], [18] and so has aspects designed to preserve
inheritability and to ensure that no part of the genome has
an inordinate effect on the phenome. With this in mind
the structure of the genome is made up from a set of
integers, rather than a binary genome.
With the data sets shown in Figs 7 and 11, C4.5 does
not produce a tree at all. Of the two data sets presented a
higher order combined attribute results in a concise tree
where no tree is produced without the higher order
attribute. In the case of the quadrant data set, a concise
decision is possible with the unaugmented data set, one is
not produced by C4.5.
x-y <= -2 : out (38.0)
x-y > -2 :
|
x-y <= 1.5 : in (52.0)
|
x-y > 1.5 : out (38.0)
A. Banded data set
This test shows a data set that does not project down
onto 1 dimension. This 2 dimensional data set results in
the following tree from C4.5, Fig. 8.
Evaluation on training data (128 items):
Before Pruning After Pruning
-------------- -----------------------Size Errors Size Errors Estimate
5
0( 0.0%) 5 0( 0.0%) ( 3.2%)
Fig. 9. The decision tree produced by C4.5 from the banded data given the
added input feature of x-y.
x <= 0.0 : (64.0)
|
y<=0.0:in(32.0)
|
y>0.0:out(32.0)
x > 0.0 : (64.0)
|
y <= 0.0 : out (32.0)
|
y > 0.0:in(32.0)
Evaluation on training data (128 items)
Before Pruning
------------------Size Errors
6 0( 0.0%)
Fig. 7. A granularised version of banded linearly separable data.
After Pruning
-----------------------Size Errors Estimate
6 0( 0.0%) ( 3.2%)
Fig. 10. The decision tree, which could be used to discriminate the
quadrant data, but cannot be produced by C4.5.
out (128.0/52.0)
Evaluation on training data (128 items):
Before Pruning
-----------------Size Errors
1 52(40.6%)
After Pruning
-----------------------Size Errors
Estimate
1
52(40.6%) (44.1%)
Fig. 8. The decision tree produced by C4.5 from the banded data.
The decision tree produced from the banded data,
Figure 8, is shown in Fig. 8 and is almost useless. It does
not reveal any useful information from the data. The most
that can be gained from this data is that there are 52
elements in the concept and the rest are out. Adding the
attribute x − y gives the tree shown in Fig. 9, this is a
good predictor and also makes the information held in the
data clear.
B. Quadrant data set
© 2012 ACADEMY PUBLISHER
Fig. 11. A granularised version of quadrant data set.
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
out (128.0/52.0)
Evaluation on training data (128 items):
function field is interpreted as 2 for plus, 3 for minus, 5
for multiply. No other function types are illustrated.
Before Pruning After Pruning
-------------- -----------------------Size Errors Size Errors Estimate
1
52(40.6%)1
52(40.6%)(44.1%)
1
Fig. 12. The decision tree produced by C4.5 from the quadrant data.
0
1
5
1
0
2
5
2
Evaluation on training data (128 items):
After Pruning
-----------------------Size Errors Estimate
5
0( 0.0%) ( 3.2%)
Fig. 13. The decision tree produced by C4.5 from the quadrant data
given the added input feature of x*y.
A. Genetic Structure
The chromosome can deliver several genes
corresponding to several combined attributes. The
chromosome is a fixed maximum length and achieves a
variable number of genes by an activation flag. Each gene
delivers one new attribute and each variable is a linear
combination of simpler variables.
1) Variable.: If the number in the variable slot is N and
there are K basic continuous variables in the data set and
M variables in the gene prior to this one then N mod (K +
M ) refers to variable within those K + M variables.
2) Function.: If the function is a monadic function then
it is applied to variable 1, otherwise to both. The
prototype system has a set of simple arithmetic functions,
power, multiplication, division and subtraction. This is
sufficient to extract all the decision trees we have
considered.
3) Number of genes and gene length.: The variable
length chromosome has disadvantages as the effect on the
gene itself of the two fields that determine the length of
the gene is considerably more than any other field and
can be destructive. The variable length gene has similar
disadvantages. The gene structure finally chosen for the
system is shown in Fig. 14.
Active
Variable 1
Function
Variable 2
Fig. 14. This shows the basic structure of the gene adopted. The
Active/Variable/Function/Variable is repeated up to the gene length.
This potentially has some of the properties of recessive
genes that are attributed to diploid gene structures
although no experiments have been conducted to
determine this. An example gene is shown in Figure 15.
This gene has 4 segments, 1 of which is active. Each
segment has 2 attributes, some active and some not. The
© 2012 ACADEMY PUBLISHER
active
x
y
inactive
x
*
x
inactive
y
*
y
inactive
x2
y2
1
3
2
x*y <= -2 : out (38.0)
x*y > -2 :
|
x*y <= 1.5 : in (52.0)
|
x*y > 1.5 : out (38.0)
Before Pruning
------------------Size Errors
5 0( 0.0%)
201
0
4
2
5
Fig. 15. An exemplar gene. x and y are variables number 1 and 2. The
first new variable is x − y and is variable number 3. As this is activated
then it will be made available as an input to the decision tree generator.
If variable 6 is activated then because it relies on variables 4 and 5 they
will also be kept but not necessarily activated.
VI. EXEMPLAR DATA
The system described above was applied to some data
sets taken from the Machine Learning Repository [19] in
order to compare the capability of the system with other
known decision tree generators.
The experiment compares the decision trees generated
by C4.5 and the decision trees generated by C4.5 with the
enhanced input space. The results consist of
• the percentage of correct results on the training set
• the percentage of correct results on the test set
• the of degrees of freedom for the decision space
• the probability that the result could not have arisen
by chance
• the decision tree size
A. Experimental results
Each data set was split randomly into two sets, the
training set which comprised 90% of the data and the test
set, which comprised 10% of the data. The split was
generated by choosing whether a particular data point
was to be in the training set or the test set using a random
number generator. This way any temporal aspects that
may be in the data are accounted for. Notice the degrees
of freedom are different for the training set and the test
set, this is because there were no data elements belonging
to one of the categories in the test set, where there were
elements in the training set.
TABLE I
EXPERIMENTAL RESULTS FOR GLASS DATA SET
Train Correct
Test Correct
DOF Train
DOF Test
probability of not null
Tree size
C4.5
92.8
75
30
25
1.0
C4.5+GP
98.5
100
30
25
1.0
43
61
202
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
The glass data set shows a considerable improvement
for the enhanced input space, however the decision tree is
larger.
The iris data set shows an improvement for the
enhanced input space, but the improvement is marginal,
however the
decision tree is smaller.
TABLE II
EXPERIMENTAL RESULTS FOR IRIS DATA SET
Train Correct
Test Correct
DOF
probability of not null
Tree size
C4.5
98
100
6
0.99
C4.5+GP
100
100
6
0.99
9
7
TABLE III
EXPERIMENTAL RESULTS FOR PIMA INDIANS DATA SET
Train Correct
Test Correct
DOF
probability of not null
Tree size
C4.5
82.7
74.5
2
1.0
C4.5+GP
98.3
84.2
2
1.0
33
119
The Pima indians data set shows a considerable
improve- ment for the enhanced input space. Both the
training set and the test set show improvement. The
enhanced decision tree is also considerably bigger, by a
factor of nearly 4.
The experiments have shown that the enhanced system
is able to significantly improve the quality of the
decisions made, however this is often at the expense of a
larger tree. The test on the iris data set indicates that the
decision tree can be smaller, as shown by some of the
demonstration data sets earlier in the paper.
VII. REDUCED DATA SETS
Work by Jensen and Shen on rough set theory aims to
reduce the input set to a subset of attributes that have the
same predictive value as the original set [20]. In this
sense whereas the work reported here extends the input
space by adding polynomials of the base features,
Jensen’s work reduces the input space. Using the glass
data set as an example set of data Jensen’s model
removes the input value that measures the amount of
Barium in the glass sample. Initial experiments show that
the reduced set does reduce accuracy, but only marginally
and not significantly, see Table IV.
TABLE IV
EXPERIMENTAL RESULTS FOR GLASS DATA SET COMPARING COMPLETE AND
REDUCED DATA SETS WITH C4.5
Train Correct
Test Correct
DOF Train
DOF Test
probability of not null
Tree size
C4.5
Complete
92.8
75
30
25
1.0
C4.5
Reduced
92.3
70
30
25
1.0
43
45
The interim conclusion is that no predictive accuracy
has been lost; but it is also true that C4.5 alone does not
extract everything from the data that it is possible to
© 2012 ACADEMY PUBLISHER
extract. The next test evolves the reduced input space to
extract as much predictive power as it can.
These preliminary tests indicate that reduction system
of Jensen and Shen [20] does not remove useful
information by eliminating input attributes, and coupled
with the enhanced input space system reported here
shows no loss of accuracy.
TABLE V
EXPERIMENTAL RESULTS FOR GLASS DATA SET COMPARING COMPLETE AND
REDUCED DATA SETS WITH C4.5 AND C4.5 + GP
Train Correct
Test Correct
DOF Train
DOF Test
probability of
not null
Tree size
C4.5
Complete
92.8
75
30
25
1.0
C4.5
Reduced
92.3
70
30
25
1.0
C4.5+GP
Complete
98.5
100
30
25
1.0
C4.5+GP
Reduced
99.0
100
30
25
1.0
43
45
61
59
VIII. CONCLUSIONS
This paper has extended the capability of decision tree
induction systems where the independent variables are
continuous. The incremental decision process has been
shown to be inadequate in explaining the structure of
several sets of data without enhancement. The paper has
shown that introducing variables based on higher order
and higher dimensional combinations of the original
variables can result in significantly better decision trees.
This can all be accomplished by introducing these
variables at the start of the decision tree generation and a
suitable method for generating these would be a genetic
algorithm. A fitness function for a genetic programming
system has been introduced and serves to discover
structure in the continuous domain. Although the work of
[20] shows how to reduce the input set without losing any
discriminating power they did not achieve all the
predictive power that the input space could provide
further work on a variety of different data sets should be
performed to confirm this.
REFERENCES
[1] J. Quinlan, “Discovering rules from large collections of
examples: a case study,” in Expert systems in the
microelectronic age, D. Michie, Ed. Edinburgh University
Press, 1979.
[2] ] J. Quinlan, “Induction of decision trees,” Machine
Learning, vol. 1, no. 1, 1986.
[3] Z. Fu, B. Golden, and S. Lele, “A GA based approach for
building accurate decision trees,” INFORMS Journal on
Computing, vol. 15, no. 5, pp. 3–23, 2003.
[4] M. Ryan and V. Rayward-Smith, “The evolution of decision
trees,” in Proceedings of the Third Annual Conference on
Genetic Programming, J. Koza, Ed. San Francisco, CA.:
Morgan Kaufmann, 1998, pp. 350– 358.
[5] R. Marmelstein and G. Lamont, “Pattern classification using
a hybrid genetic program-decision tree approach,” in
Proceedings of the Third Annual Conference on Genetic
Programming, J. Koza, Ed. San Francisco, CA 94104,
USA: Morgan Kaufmann, 1998.
[6] G. Folino, C. Puzzuti, and G. Spezzano, “Genetic
programming and simulated annealing: a hybrid method
to evolve decision trees,” in Pro- ceedings of the Third
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 4, NO. 2, MAY 2012
European Conference on Genetic Programming, R. Poli,
W. Banzhaf, W. Langdon, J. Miler, P. Nordin, and T.
Fogarty, Eds. Edinburgh, Scotland, UK: Springer-Verlag,
2000, pp. 294–303.
[7] D. Greene and S. Smith, “Competition-based induction of
decision models from examples,” Machine Learning, vol.
13, pp. 229–257, 1993.
[8] A. Freitas, “A GA for generalised rule induction,” in
Advances in Soft Computing, Engineering Design and
Manufacturing. Berlin: Springer, 1999, pp. 340–353.
[9] D. Carvalho and A. Freitas, “A genetic-algorithm based
solution for the problem of small conjuncts,” in Principles
of Data Mining and Knowl- edge Discovery (Proc. 4th
European Conf., PKDD-2000. Lyon France), ser. Lecture
Notes in Artificial Intelligence, vol. 1910. SpringerVerlag, 2000, pp. 345–352.
[10] K. De Jong, W. Spears, and D. Gordon, “Using a genetic
algorithm for concept learning,” Machine Learning, vol.
13, pp. 161–188, 1993.
[11] C. Janikow, “A knowledge intensive GA for supervised
learning,” Machine Learning, vol. 13, pp. 189–228, 1993.
[12] Y. Hu, “A genetic programming approach to constructive
induction,” in Genetic Programming, Proceedings 3rd
Annual Conference, San Mateo, California, 1998, pp.
146–151.
[13] A. Papagelis and D. Kalles, “GA tree: Genetically evolved
decision trees,” in Tools with AI, ICTAI Proceedings 12th
IEEE International Conference, 13–15 Nov 2000, pp.
203–206.
[14] J. Eggermont, J. Kok, and W. Kosters, “Genetic
programming for data classification: Partitioning the
search space,” in ACM Symposium on Applied
Computing, 2004, pp. 1001–1005.
[15] D.Michie and R.Chambers,“Boxes:an experiment in
adaptive control,” in Machine Intelligence 2, E. Dale and
D. Michie, Eds. Edinburgh: Oliver and Boyd, 1968.
[16] A. Konstam, “Linear discriminant analysis using GA,” in
Proceedings Symposium on Applied Computing,
Indiannapolis.IN, 1993.
[17] M. Withall, “Evolution of complete software systems,”
Ph.D. disserta- tion, Computer Science, Loughborough
University, Loughborough, UK, 2003.
[18] M. Withall, C. Hinde, and R. Stone, “An improved
representation for evolving programs,” Genetic
Programming and Evolvable Machines, vol. 10, no. 1, pp.
37–70, 2009.
[19]
“Machine
learning
repository,”
http://archive.ics.uci.edu/ml/datasets.html, 2010.
[20] R. Jensen and Q. Shen, “New approaches to fuzzy-rough
feature selection.” IEEE Transactions on Fuzzy Systems,
vol. 17, no. 4, pp. 824–838.
Chris J. Hinde is Professor of
Computational Intelligence in the
Department of Computer Science at
Loughborough University. His interests
are in various areas of Computational
Intelligence including fuzzy systems,
evolutionary computing, neural networks
and data mining. In particular he has
been working on contradictory and
inconsistent logics with a view to using them for data mining. A
recently completed project was concerned with railway
scheduling using an evolutionary system. He has been funded
by various research bodies, including EPSRC, for most of his
career and is a member of the EPSRC peer review college.
Amongst other activities he has examined over 100 PhDs.
© 2012 ACADEMY PUBLISHER
203
Anoud I. Bani-Hani is an EngD research
Student at Lough- borough University,
researching into Knowledge Management in
SMEs in the UK, with specific focus on
implementing an ERP system into a lowtech SME. Prior to joining the EngD
scheme Anoud was a Lecturer at Jordan
University of Science and Technology and
holds an undergraduate degree in Computer
science and information technology system from the same
university and a Master degree in Multimedia and Internet
computing from Loughborough University.
Thomas W. Jackson - is a Senior
Lecturer in the Department of Information
Science at Loughborough University.
Nicknamed ‘Dr. Email’ by the media Tom
and his research team work in two main
research areas, Electronic Communication and Information Retrieval within
the Workplace, and Applied and Theory
based Knowledge Management. He has
published more than 70 papers in peer reviewed journals and
conferences. He is on a number of editorial boards for
international journals and reviews for many more. He has given
a number of keynote talks throughout the world. In both
research fields Tom has, and continues to work closely with
both private and public sector organisations throughout the
world and over the last few years he has brought in over £1M in
research funding from research councils, including EPSRC. He
is currently working on many research projects, including ones
with The National Archives and the Welsh Assembly
Government surrounding information management issues.
Yen P. Cheung has an honours degree in
Data Processing from Loughborough
University of Technology (UK) in 1986
and completed her doctorate in
Engineering
at
the
Warwick
Manufacturing Group (WMG), University
of Warwick in 1991. She worked initially
as a Teaching Fellow and then as a Senior
Teaching Fellow at WMG from 1988
1994 where she was responsible for the AI and IT courses in the
M.Sc. programs in Engineering Business Management. She also
ran IT courses for major companies such as British Aerospace,
Royal Ordnance, London Electricity and British Airways in UK
and for Universiti Technology Malaysia in Malaysia. Besides
teaching, she also supervised a large number of industry based
projects. After moving to Australia in 1994, she joined the
former School of Business Systems at Monash University.
Currently she is a Senior Lecturer at the Clayton School of IT at
Monash University, Australia where she developed and
delivered subjects in the area of business information systems,
systems development, process design, modelling and
simulation. Her current research interests are in the areas of
collaborative networks such as eMarketplaces particularly for
SMEs, intelligent algorithms for business systems, applications
and data mining of social media. She publishes regularly in both
international conferences and journals in these areas of research.
Call for Papers and Special Issues
Aims and Scope
Journal of Emerging Technologies in Web Intelligence (JETWI, ISSN 1798-0461) is a peer reviewed and indexed international journal, aims at
gathering the latest advances of various topics in web intelligence and reporting how organizations can gain competitive advantages by applying the
different emergent techniques in the real-world scenarios. Papers and studies which couple the intelligence techniques and theories with specific web
technology problems are mainly targeted. Survey and tutorial articles that emphasize the research and application of web intelligence in a particular
domain are also welcomed. These areas include, but are not limited to, the following:
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Web 3.0
Enterprise Mashup
Ambient Intelligence (AmI)
Situational Applications
Emerging Web-based Systems
Ambient Awareness
Ambient and Ubiquitous Learning
Ambient Assisted Living
Telepresence
Lifelong Integrated Learning
Smart Environments
Web 2.0 and Social intelligence
Context Aware Ubiquitous Computing
Intelligent Brokers and Mediators
Web Mining and Farming
Wisdom Web
Web Security
Web Information Filtering and Access Control Models
Web Services and Semantic Web
Human-Web Interaction
Web Technologies and Protocols
Web Agents and Agent-based Systems
Agent Self-organization, Learning, and Adaptation
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Agent-based Knowledge Discovery
Agent-mediated Markets
Knowledge Grid and Grid intelligence
Knowledge Management, Networks, and Communities
Agent Infrastructure and Architecture
Agent-mediated Markets
Cooperative Problem Solving
Distributed Intelligence and Emergent Behavior
Information Ecology
Mediators and Middlewares
Granular Computing for the Web
Ontology Engineering
Personalization Techniques
Semantic Web
Web based Support Systems
Web based Information Retrieval Support Systems
Web Services, Services Discovery & Composition
Ubiquitous Imaging and Multimedia
Wearable, Wireless and Mobile e-interfacing
E-Applications
Cloud Computing
Web-Oriented Architectrues
Special Issue Guidelines
Special issues feature specifically aimed and targeted topics of interest contributed by authors responding to a particular Call for Papers or by
invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are of interest to the Journal.
Preference will be given to proposals that cover some unique aspect of the technology and ones that include subjects that are timely and useful to the
readers of the Journal. A Special Issue is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length.
The following information should be included as part of the proposal:
•
Proposed title for the Special Issue
•
Description of the topic area to be focused upon and justification
•
Review process for the selection and rejection of papers.
•
Name, contact, position, affiliation, and biography of the Guest Editor(s)
•
List of potential reviewers
•
Potential authors to the issue
•
Tentative time-table for the call for papers and reviews
If a proposal is accepted, the guest editor will be responsible for:
•
Preparing the “Call for Papers” to be included on the Journal’s Web site.
•
Distribution of the Call for Papers broadly to various mailing lists and sites.
•
Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be
informed the Instructions for Authors.
•
Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact
information.
•
Writing a one- or two-page introductory editorial to be published in the Special Issue.
Special Issue for a Conference/Workshop
A special issue for a Conference/Workshop is usually released in association with the committee members of the Conference/Workshop like general
chairs and/or program chairs who are appointed as the Guest Editors of the Special Issue. Special Issue for a Conference/Workshop is typically made of
10 to 15 papers, with each paper 8 to 12 pages of length.
Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop:
•
Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers of XYZ Conference”.
•
Sending us a formal “Letter of Intent” for the Special Issue.
•
Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees.
Information about the Journal and Academy Publisher can be included in the Call for Papers.
•
Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus the
evaluation from the Session Chairs and the feedback from the Conference attendees.
•
Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors
should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced.
•
Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact
information.
•
Writing a one- or two-page introductory editorial to be published in the Special Issue.
More information is available on the web site at http://www.academypublisher.com/jetwi/.
(Contents Continued from Back Cover)
The Developing Economies’ Banks Branches Operational Strategy in the Era of E-Banking: The Case
of Jordan
Yazan Khalid Abed-Allah Migdadi
189
Evolving Polynomials of the Inputs for Decision Tree Building
Chris J. Hinde, Anoud I. Bani-Hani, Thomas.W. Jackson, and Yen P. Cheung
198