Large Scale Evaluation of an Adaptive E-Advertising
User Model
Alaa A. Qaffas and Alexandra I. Cristea
Department of Computer Science, The University of Warwick, Coventry, CV4 9LX, UK
{aqaffas, acristea}@dcs.warwick.ac.uk
Abstract. The user model is considered the core content of any adaptive system.
The process that is most challenging in an adaptation system is the modelling of
user data. This paper is focused on creating a user model that can be suitably
adapted to advertisement applications. To integrate data from different publicly
available sources, a social component has been added to the user model, to support advertisements adaptation. In addition, a straightforward and lightweight
user model has been targeted, which should be easy to integrate into any existing
system. A tool based on this model is introduced and a study that assesses the
effectiveness of it via a trial run of a prototype of the tool with different users is
described.
Keywords: E-advertising, E-commerce, Personalisation, Adaptive Advertising,
User Model, User Profile.
1
Introduction
Technological advancements have led to a significant increase in web-based promotions and online marketing, as target audiences can now be accessed regardless of time
or location. The adaptation of advertising adds significant benefits to customer satisfaction and business profits [1]. Availability and the easiest way to manage the adaptation of advertising with minimal effort on any commercial site has become a key demand for businesses. Many models exist, e.g., the Dexter model [2], AHAM [3], and
LAOS [4], but they are proposed mainly for personalising the educational experience.
Lessons learned from them may be applied here, to some extent. Moreover, these models do not feature the lightweight integration of adaptive features on any website as
their main purpose. Therefore, a new model - the Layered Adaptive Advertising Integration – has been proposed, based on prior ones, in order to introduce an easier approach to integrating adaptation features into any commercial website. Some of its components differ from those in traditional models. In particular, the user modelling in the
proposed model is separated into storage and delivery parts, the latter is not covered in
this paper, the latter having been implemented, but data is still being analysed. This
separation can potentially enhance the generalisation, portability and efficiency of the
user model and delivery model. The storage part is encapsulated and manipulated via
XML representation, to allow the system to be integrated into any website easily and
adfa, p. 1, 2011.
© Springer-Verlag Berlin Heidelberg 2011
with only minor changes to the original database of the website. In addition to the separation, the storage of the user model and its operation is added to the delivery model,
in order to facilitate the integration process on any website. This separation also allows
the new model to be easily expanded.
Our research aims to address the following main research question:
How can we support website owners in the creation of adaptive advertising?
This main research question can be addressed by answering the following sub-research questions:
A. What type of tools do website owners need, to be able to efficiently add adaptive
advertising in a lightweight manner (as an add-on) to their website?
B. What kind of support do website owners need, to be able to use these tools?
To answer these questions, we recommend a collection of tools, Adaptive E-Advertising Delivery System (AEADS), which facilitate the creation of adaptive e-advertising. This paper in particular focuses on one of the vital components in any adaptive
systems, the User Model (UM). In this paper, we propose a lightweight UM, with a set
of features and attributes that we consider essential to adaptive advertising, and which
can be easily added to any static commercial website. Furthermore, this model is implemented and evaluated with real Internet users and customers.
The following sections discuss the related research, the user model tool and its evaluation, and finally provide a conclusion.
2
Related Research
Adaptive hypermedia systems allow for personalisation, thus improving the efficiency
and accuracy of information distribution [5]. This process consists of three major types
of tasks: acquisition, representation and secondary inference, and production [6]. The
acquisition tasks identify information regarding users’ characteristics, computer usage
and environment, in order to construct an initial model of the user. The representation
and secondary inference tasks inference and express the content of the user model and
makes assumptions about them, such as their behaviours and the environment. The production tasks generate the adaptation of the contents and structure of the system to meet
the users’ needs. We use the classification of the data in a user model by user, usage,
environment, hardware, software, and location. According to [7], an adaptive model
considers user’s characteristics in developing a user model and the data captured can
be categorised as knowledge, interests, goals and tasks, background, individual traits,
and context of work.
A user model is a basic component in any personalised system and is a representation
of user data stored for any adaptive changes to the system's behaviour. All adaptive
hypermedia frameworks and models have a user model as one of their components. For
instance, in AHAM [3], the user model contains concepts with attributes storing user
preferences, while in LAOS [4] the user model is even more complex.
There have been many systems proposed to facilitate adaptation, including AHA!
[8, 9], GALE [10], ADE [11], AdROSA [12], ADWIZ [13], MyAds [14], and
WHURLE [15]. A generic user model based on variable-value pairs that facilitate required adaptations form the basis of ADE, AHA! and GALE and this is suitable for
online advertising purposes as well. In particular, according to [16] a model is suggested for the delivery of hypermedia content that considers the types of users, the devices used by customers to gain access, the types of access, the state of the network and
the current load on the server. Nonetheless, these are all standalone systems, which
cannot be integrated into existing ones in a lightweight manner. XML-based pipelining,
as used by WHURLE for applying lightweight solutions and standards, is efficient for
adding minor modifications to existing systems, and is, therefore, utilised in our approach. However, user modelling in WHURLE is not as extended.
AdSense [17, 18], unlike our approach, cannot provide advertisements to clients directly, it just lets advertisers in the Google Network deliver advertisements to the content site to be presented to users automatically. It specialises in banner advertisements
and uses location to personalise content [19]. However, this process does not utilise any
form of user-based modelling, or the assimilation of user information for personalisation purposes. Among the potential approaches to selecting the best form of advertisement, adaptive hypermedia may be used to link the advertisement to the consumer's
taste, via user modelling, and is a significant element within systems that adapt to the
user [20].
Another example, AdROSA [12], that makes automatic personalised web banners,
depends mostly on specific browsing behaviours of a user. It is similar to AdSense, the
portal model of advertising uses AdROSA to deliver the advertisements. More recent
systems emerged, such as ADWIZ, and MyAds, which are adaptive advertisement systems. ADWIZ concentrates on the adaptation of web banners and is mainly dependent
upon keywords, supplied via a search-service query, in order to select advertisements.
Furthermore, this system does not require the use of cookies or, indeed, any user-related
information within its database, with which to identify a particular user. The MyAds
system concentrates on those advertisements within a public-display scenario by dynamically adapting them for an audience within a certain proximity of the display. A
more recent system also called MyAds [21] is a social adaptive hypermedia system
used for online advertising. It is a standalone system based on its own theoretical framework, build as five main layers. The theoretical framework is rooted in the adaptive
hypermedia theory.
Social networks are good sources of user information [22], from which user behaviour and characteristics for personalising advertisements can be retrieved. Although the
type of content posted on these websites varies, it is generally indicative of a user’s
preferences, attitudes and behaviours.
Facebook is one of the most popular social networking sites, with 1.35 billion
monthly active users from around the world [23]. Users can create a personal profile,
add friends, send messages, post status updates and comments to friends’ “walls”. They
can chat together and upload photos and videos that their friends can comment on and
“like” [24]. For these reasons, Facebook has been used as the first social medial datagathering source for the first version of the system described in this paper, follow-up
versions look into other sources though.
Many existing semantic web-authoring systems can be used in conjunction with
other delivery or authoring systems [25, 26]. In our case, XML was selected to generate
the user model tool’s internal representation.
The modelling of user profiles involves acquisition, representation and secondary
inference. Data acquisition can be performed using a variety of different methods depending on class, including user data acquisition methods, usage data acquisition methods and environment data acquisition methods. This includes user-supplied information
acquired through questions asked by the system, acquisition rules, stereotype reasoning,
and plan recognition, a process which predicts future actions based on previous patterns
[27]. A simple method for making a first assessment of others is to classify them into
groups sharing the same interests, according to a set of criteria – a stereotype [28]. We
use the stereotype technique, as it makes inferences based on limited observations.
3
Authoring Adaptive E-Advertising
The overall Authoring model of Adaptive E-Advertising, as informed by prior research
and implementations, especially in the area of personalised e-learning, includes:
1. The Domain Model - used by businesses to organise, label and categorise advertisements. As it has been described elsewhere [29], it is not further detailed
here.
2. The Adaptation Model [30] - enabling businesses to adapt the advertisements
they have organised, using the domain model tool for their customers’ needs.
3. The User Model - representing the personal data of an individual user, stored
for any adaptive changes to system's behaviours. For example, it can be used
to predict the most relevant items for the user, when they search for information, as described below. This is the focus of our paper.
Here, the social input data component has been added to the user model, and then
some functions of this model were separated, (e.g., the inference function) to be used
in the delivery engine to support the integration process.
The user (customer) modelling tool has been designed to be simple (to have few user
model features), in order to be lightweight, and to integrate with any potential website
user model. With this tool, we implement the first steps of the user modelling, including
its acquisition data, and retrieve explicit and implicit data. We use the explicit data
supplied by users, and retrieve data from social networks, by using the social networks
authorisation, and authentication APIs. We also conduct implicit data acquisition, by
using several techniques, including stereotype reasoning [31], and plan recognition
[27] to be used in the delivery Engine.
All of the data about users in the user model is stored in XML files. Storing all of
the data in a lightweight fashion (XML) facilitates the integration into any commercial
webpage, as XML allows for pipeline processing and independence to any other processing on the website. Users can login into the system via two methods: register (Fig.
1), and Facebook login (Fig. 2). By logging in via the latter, the user model can be
automatically populated with the necessary information for the adaptation of advertisements. The user information is arranged in an XML file with attributes such as id,
name, password, email, age, gender, location, number of
logins to site, total number of clicks on each advertisement, device
used, and software used. All users can update their information on their profile page.
This data will be stored in the users.xml file (Fig. 3). The social data from the social
login allows us to retrieve sufficient information about users and to infer from specific
to general cases.
Fig. 1. User Registration.
Fig. 2. User Login.
Fig. 3. User Model XML file sample.
The implementation of the user modelling is made by creating servlets to be used in
JSP pages, adding data to the user_item.xml file, such as the number of clicks on
advertisement for each user, and the number of times each advertisement is shown, for
each user attribute (Fig. 4 a, b and c). The number of clicks and shows will be utilised
to apply plan recognition. Plan recognition refers to the task of inferring the plan of an
intelligent agent (here, the human customer) from observing the agent's actions or their
effects [27]. In addition, this process will depend on the plan library that businesses
create in the authoring part. The delivery engine checks the clicked items and the plan
library to acquire a sequence of advertisements to be presented to the user. The latter
process belongs to the delivery engine part and no further details are given here. Moreover, a new XML file named users_items_sequence.xml tracks each user’s
selection sequence of advertisements, albeit only the final ten selections will be stored
in this file. The threshold of 10 selections was decided based on trial and error on the
testing phase of the system. This file will be used to predict user actions for current and
similar users.
a. Show Item
b. Item Details
c. User Item.XML file
Fig. 4. User Model Instantiation Example.
Many formats can be used in a user model, for example attribute-value pairs, Booleans,
lists, references to external objects, and so on. The knowledge in a user model can be
represented in different ways, such as overlay models, semantic nets, user profiles and
stereotype-based models. For example, in an overlay model the user’s knowledge is
represented as a subset of the domain model of the application.
Furthermore, for advertising adaptations, we used, as said, the stereotype technique,
as it makes inferences based on limited observations. Each user is assigned to a group
(stereotype), according to the types of advertisements on websites (a website owner
arranges his advertisements into groups and subgroups). The system then determines
the activation conditions for applying the stereotype to a user. For example, if the user
model shows that the user is interested in computers and televisions, then the system
activates the stereotype “technology”. From the usage data, if, for instance, the user has
bought at least two electronic items or computers, then the stereotype “technology” can
be activated. The administrator can create and control (add- update - delete) stereotypes.xml from the stereotype page (Fig. 5). Based on hypotheses H1 (described below), an initial minimal set of necessary dimensions for an advertising user model are
defined, and include age, gender, bandwidth, device type, number of clicks on advertisements, education level, education type, and hobbies. Each dimension has its own
attributes. In addition, action sequencing is used in this research to predict the future
actions of the user, to recommend actions based on the action sequences of other users,
or to perform some of these actions on behalf of the user.
Fig. 5. Stereotype.
4
Scenario
To better understand the usage of the user modelling tool and the application of the data
it stores, we describe two usage scenarios as follows:
4.1
Scenario 1
When the login page is loaded, Albert, a 20-year-old man, and a customer of a given
company, enters his username and password. He could click the login button or he can
login using his Facebook account. In both cases, bandwidth, location, device type, and
software used for Albert are automatically obtained by the system. Login with a Facebook account will simplify access to the website and allows systems to automatically
obtain important data.
Albert is using a smart phone with bandwidth lower than 1M (as extracted by the
system). If Albert logs into the website for the first time, then only the general rules will
be applied. All of the advertisements that are not appropriate for Albert (based on general rules: e.g., advertisements targeted to women, to higher or lower ages, higher
bandwidth, or another device type) will be excluded. All of the advertisements that are
appropriate for Albert, and all of the advertisements that are not assigned any rules,
will be placed in the queue, to be shown to Albert.
However, if Albert logs into the website more than once, the behaviour rules and
some inference processes will be applied. In order to apply behaviour rules and inference processes, the system needs to store all of Albert’s behaviour, the advertisements
that are shown to him and not clicked, the numbers of his actions, as well as advertisements that were shown and clicked, and the number of times they were clicked. When
Albert clicks on any advertisement link, or the advertisement is shown to him, the system stores all of this data in two fields (number of shows, and number of clicks for each
advertisement). In addition, the system stores the last ten clicks for Albert to be used to
infer his actions.
Based on this data, the system applies the behaviour rules and places the advertisements that results from it into another queue; in addition, there is another queue for
the inference process, based on Albert’s history of actions. Finally, the system decides
on an advertisement (or collection thereof) from these three queues to be shown on the
page that Albert loaded.
4.2
Scenario 2
John, a 30-year-old man, logs into his favourite real estate website that displays an
array of different advertisements including real estate, video games and movie advertisements. John is uninterested in video games and prefers not to see video game advertisements. The business that owns the website can add—via the AEADS user model
and its tools—a 'like' or 'stop' button for every advertisement featured on the website.
Consequently, if John presses a ‘stop’ button for a video game advertisement, then, on
the basis of the rules of the business, the inference engine is able to stop the same advertisement, or indeed, all advertisements that belong to the same category.
Conversely, if John presses the ‘like’ button for any given movie advertisement, the
inference engine will display this advertisement a few more times, according to, and
determined by, his business preferences, and will also display a number of advertisements located within the same group as the advertisement. The system will then automatically infer that John likes items in the same group and mark them with ‘like’. John's
activity can then be stored within the social component of the user model and, based
on business rules, his preferences will prevail through follow-up log-ins.
5
Case Study
5.1
Hypotheses
The following hypotheses have been defined to evaluate the user model tool:
H0a: The user model (UM) concept for advertising (as illustrated by the UM tool)
is useful for constructing a user model for recommendation of advertisements.
H0b: The UM concept for advertising (as illustrated by the UM tool) is easy to
use for constructing a user model for recommendation of advertisements.
H0x are the basic hypotheses, and the rest are derived from them:
H1: The attributes of the proposed UM are useful for recommending advertisements (username, password, email, age, gender, education level, education
type, hobbies, bandwidth, location, device type and software used).
H2: The data in the user model is adequate for the advertisements delivery engine
decision.
H3: Automatically generating user model data (location, device type, and software used) is useful.
H4: Social networks used as a source for user data are an appropriate data
source for recommending advertisements.
H5: A user’s advertisement preferences can be predicted, by tracking the user’s
behaviour sequence when they use the system.
H6a: The input and output mechanisms of the user model tool are useful.
H6b: The input and output mechanisms of the user model tool are easy to use.
H7: The stereotypes for users with respect to advertisements recommendation are
useful and appropriate.
H8: The stereotypes for users with respect to advertisements recommendation are
easy to use.
H9: It is useful to integrate the user model creation tool in any JSP website.
H10: It is easy to integrate the user model creation tool in any JSP website.
H11: Any website administrator can understand, use, and update the stereotypes.
These hypotheses were evaluated by surveying a sample group of Internet users and
analysing their answers, as further described below.
5.2
Case Study Setup
The user model tool was evaluated from a functionality and ease of use perspective by
students studying different subjects and modules (Introduction to Business, Principles
of Marketing, Management Information System and E-Marketing) at King Abdul-Aziz
University in Jeddah, Saudi Arabia. Students were deemed appropriate as a testing population because, first, all of them are Internet users, and regular online shoppers, who
are familiar with the current online providers. The other reason was to get a large number of users. Note that, whilst our users were familiar with the Internet, their study of a
variety of subjects ensured that they were not only Computer Science specialists, and
that the tool was tested with a wide variety of backgrounds, knowledge and interests.
Consequently, a sample of 285 Internet users were asked to evaluate the user model
tool. In assessing the tool, they were asked to do the following.
First, the respondents were introduced to the user model tool and given a general
overview of adaptive advertising. Next, the participants were instructed to use the tool
and assess its effectiveness. The three-part questionnaire was provided at this point, to
guide the evaluation process. The first section collects data on the personal details of
each user. The second part presents a series of Likert scale [32] questions, to encourage
the users to rate the effectiveness of the system in terms of functionality and application.
The Likert scale offered each respondent a series of five options when evaluating the
user model tool, with the first scale option being ‘not at all useful’ or ‘very difficult’
and the last scale option being ‘very useful’ or ‘very easy to use’, respectively. A series
of qualitative questions were posed in the final section, for respondents to speak freely
about their experiences using the user model tool.
Table 1. Authoring Tool Features.
A
B
C
D
E
F
Whole User model Tool
User Registration Process
Login Process
Facebook Login Process
Submitting Information
Updating User Profile
G
H
I
J
K
L
Saving Information in XML as Export Format
Facebook User Profile Import
Match User Characteristic with Stereotype
Adding own Stereotype
Modifying existing Stereotype
Deleting Stereotype
Table 2. User Model Attributes.
1
2
3
4
5
6
7
8
9
5.3
Location
Device Type
Software Used on Device
Username
Passwords
Email
Age
Gender
Education Level
10
11
12
13
14
15
16
17
18
Education Type
Hobbies
Bandwidth
Get Location Automatically
Get Device Type Automatically
Get Software Used Automatically
Getting Number of Shows for Each User
Getting Number of Clicks for Each User
Getting Last 10 Sequence of Clicks for Each User
Results
A number of 134 survey questionnaires of the total 305 questionnaires that had been
distributed were returned completed. The reason why less than half of the distributed
questionnaires were completed may have been the fact that, prior to the survey, the
participants were informed that it was not compulsory for them to answer the questionnaire and that if they did not, their academic activities or outcomes would not be affected in the least. The majority of the participants who completed the questionnaire
were in the age group 18-24 years (61.2%), whereas most of the rest were in the age
group 25-34 years (38.1%). Furthermore, the proportion of male participants was
70.9% while the proportion of female participants was 29.1%. Moreover, results reveals
that the bachelor’s degree was the level of education of the majority of the participants,
although a small percentage of participants (11.2%) were educated to postgraduate
level. Due to these participant statistics, the research data may have been somewhat
biased in favour of young adults with a good level of education. Nonetheless, the data
are not erroneous, as this section of the demographic is the one that is dominating current and future demand, and therefore should be prioritised by web developers.
In order to make it as easy as possible for the participants to evaluate the features
and functions of the tool, the Likert scale (Figure 6) has been employed. Results showed
that there were positive reactions to every main feature (A-L, as indicated in Table 1).
Most participants gave the features a rating of no less than four, with a standard deviation of 0.46-0.51 and mean value of 4.47-4.69; this suggests that the features of the tool
were considered by the participants to be practical. Overall, then, the user model tool
can be classified as ‘useful’, because all scores were greater than three. In addition, the
Cronbach’s Alpha score is 0.91 [≥ 0.9], meaning that the reliability of the psychometric test is excellent [33]. However, despite the general positive welcome, some features
proved to be more popular than others with the participants. Thus, ‘Saving Information
in XML as Export Format’ and the ‘Facebook Login Process’ were the features with
the highest level of overall popularity, whilst on the other hand, ‘Match User Characteristic with Stereotype’ and ‘Modifying existing Stereotype’ were the features that enjoyed less popularity. This may have been due to the fact that the intended function of
these features was not understood clearly by the participants (especially terms such as
'stereotype'). Nevertheless, the less popular features are still considered very useful,
because all of them received a score higher than four. Consequently, hypotheses H6a
and H7 are validated by these results, highlighting not only the usefulness of the input
and output mechanisms of the user model tool, but also the usefulness and relevance of
user stereotypes associated with advertisements suggestion.
Fig. 6. Usefulness for Features (Ox axis detailed in Table 1).
There was wide agreement among the participants with regard to the idea that, in
order to select suitable advertisements that are compatible with their profile and preferences, every user model attribute should be collected. As can be seen in Figure 7, the
participants classified the attributes of the user model as either useful or very useful,
which has been confirmed by their mean score of 4.46-4.68 and the standard deviation
of 0.47-0.51. In addition, the Cronbach’s Alpha score is 0.92  [≥ 0.9], meaning that
the reliability of the psychometric test is excellent [33]. However, two attributes of the
user model were considered to be more useful than the others, and those were “Education Level” and “Bandwidth”, both of which received extremely high ratings. To a certain degree, therefore, hypotheses H1and H2 are confirmed by these results, as they
argued in favour of the utility of the attributes of the proposed user model for advertisement suggestion. On the other hand, there were two attributes that were associated
with notably low scores, and these attributes were ‘Get Software Used Automatically’
and ‘Getting Last 10 Sequences of Clicks for Each User’. In spite of scoring low, these
two attributes still had scores higher than four. Furthermore, one potential explanation
as to why these attributes were unpopular among the participants may be the fact that
the participants were made anxious by the idea of their behaviour and activities being
monitored. As a result, the attributes were not deemed to be of paramount importance,
although their utility was nevertheless acknowledged. Thus, hypotheses H2 (computed
from all user model attributes), H3 (computed from generated user model attributes:
13-16 in Table 2) are validated by these results. H5 needs to be analysed from tracking
data, from the sequence of clicking on advertisements. Monitoring of the order in which
users carry out their activities when making use of the system is an effective way of
anticipating what kind of advertisements the users prefer [34]. Moreover, engine decisions with regard to advertisement recommendations benefit from the collected user
model data, as well as from automatic production of user model data.
The average for all the user model attributes is of 4.55. When compared with the
neutral response (of 3), this shows a difference of 1.55.
Performing a paired T-test for all users, comparing their average score for all user
attributes, with the neutral response, the T-value is of 99.29, and the probability is of
0.0001 < 0.05 (the significance threshold most commonly used in significance research).
This result shows that the user model attributes are appreciated by the users in our
test sample, and that the positive difference, when compared to a neutral response of 3
is statistically significant.
Fig. 7. Usefulness for UM attributes (Ox axis detailed in Table 2).
The results obtained from the survey questionnaire also revealed the fact that the
participants were of the opinion that no feature of the proposed user model presented
any difficulty of use. This was confirmed by the mean scores between 4.46-4.63 attained for this dimension, as well as by the standard deviation of 0.48-0.50. Moreover,
Cronbach’s Alpha for ease of use is 0.90, which means that the level of reliability is
excellent [33]. What is more, the findings of the data analysis indicated that some features were better received by the participants than others. Thus, the features of ‘User
Registration Process’ and ‘Updating User Profile’ enjoyed a highly favourable response
from the participants; on the other hand, the features of ‘Adding own Stereotype’ and
‘Modifying existing Stereotype’ were less favourably looked upon by the participants.
Hence, these findings serve to confirm hypotheses H7 and H8, which put forth the argument that the advertisement recommendations were suitable and did not pose any
obstacle to use. It must be noted that, despite being less well-received by the participants, the latter two features still scored above four, meaning that they are not difficult
to use. On the whole, the ease of use of the user model tool is supported by the results
of the analysis of the survey data.
Fig. 8. Ease of Use for Features (Ox axis detailed in Table 1).
5.4
User Tracking
The survey questionnaire explains that during tracking user’s actions, the number of
clicks is increased with increasing the time of system usage. This information can
reflects the predilections of the use for our system with time progress. Thus, we can
conclude that the advertisements become more appropriate for users as the tracking of
user’s action has been applied, as illustrated in Figure 9. Overall, the data in the user
model that is collected from users’ tracking supports the possibility that such a system
attracts users to view advertisements. This somewhat supports out hypothesis H5, that
tracking of the users can illustrate their advertisement preferences.
Fig. 9. Clicks progress against time.
The users’ tracking data shows that the advertisements in the category books have a
higher rate of clicks, as shown in Figure 10. Businesses categorise the advertisements
in the first level of adaptation based on user’s characteristics. Thus, according to the
dominant characteristics of participants, the age group of 18-24 years, and the education
level of bachelor’s degree for most of the participants, the book group become the most
highly clicked on one by participants.
Fig. 10. Number of clicks for different groups.
We also track sub-categories, such as books sub-groups. From these, the most popular are computer science books as demonstrated in Figure 11.
Fig. 11. Number of clicks for books sub-groups
5.5
Participant Feedback
In addition to the questions that were designed to shed light on the participants’ views
with regard to the various features and functions of the user model tool, the survey
questionnaire also included a section in which the participants were requested to provide feedback about the tool, so as to help identify the dimensions that required improvement. Such feedback was considered to be an essential part of the study, due to
the fact that it assisted in the resolution of any emerging problems with the system, and
therefore it aided in increasing the performance of the user model tool. Thus, important
insight and recommendations were derived from the feedback provided by the participants. As remarked by a number of participants, the expansion of the scope of targeted
campaigns could be achieved through the diversification of the spectrum of hobbies
and leisure activities supplied by the tool in the form of features. Meanwhile, other
participants expressed their satisfaction with the feature for Facebook login, which most
web-authoring applications have incorporated. The great usefulness of such a feature
resides in the fact that it not only supports the user model becoming better integrated
with other web-based systems, but it also makes the model more functional and easy to
use, since users can gain access to a range of different applications by entering their
identification details only a single time. To some degree, these findings are in line with
hypothesis H4, which maintained that advertisement recommendations could draw on
the source of user data that is supplied by platforms of social networking. An additional
suggestion of relevance that was made by one participant was that the model tool should
not specify age in letters but in numbers. On the other hand, no remarks with regard to
potential improvements that could be brought to the tool were made by some participants, although they did express their agreement that such research is necessary and
significant, particularly as online marketing systems and tactics are constantly being
innovated and transformed. On the basis of such feedback, it can be inferred that the
user model tool should be amended so as to become not so much a highly technical
process that cannot be used without a certain level of specialist knowledge, but more of
a mechanism that is easily accessible to users in all aspects. Meanwhile, calculation of
bandwidth was reported by some participants as confusing and unclear; therefore, in
order to be able to better monitor bandwidth for the users, bandwidth restrictions should
be automatically taken into account. In addition, the storage of data in XML rather than
in a database was questioned by one participant, which raised awareness about the necessity to provide a clear explanation as to the manner in which the transfer of XML
data between various programmes can be achieved without any difficulty. To some
extent, these results serve to validate hypotheses H6a and H6b, which argued in favour
of the usefulness and ease of use of the input and output processes of the user model
tool. Moreover, the system can be easily run by all users, regardless of the level of
system knowledge they possess.
The feedback provided by one participant addressed the fact that the creation of more
comprehensive and detailed user profiles and the identification of particular target
groups of users require the collection of greater amounts of demographic data. Along
similar lines, a different participant recommended that the tool should be diversified
through the introduction of a larger range of dimensions and the formulation of more
clear-cut rules with the purpose of devising marketing strategies of higher efficiency.
However, despite the fact that all the feedback provided by the participants was relevant
and helpful, care should be exercised when taking these recommendations on board,
since the aim was the creation of a user model characterised by flexibility, ease of use,
applicability, and transferability. Careful implementation of suggestions is essential, as
previous experience with adaptive hypermedia has proven that confusion and lack of
clarity may be exacerbated rather than diminished by the addition of more features. One
solution to this that was put forth within the setting of adaptive education [35] is to
enable users to customise and adapt the different features in accordance with their needs
as well as their experience and understanding of adaptive processes and systems. In
addition, it is important to remember that the purpose of this study is not to formulate
related strategies but to create a user model. Last but not least, with respect to interface
and usability, one participant observed that the user interface of the system needs to be
made more appealing and attractive; it cannot be doubted that improving the aesthetic
appearance of the user interface would have a positive impact on users’ attitudes and
responses, but it does not represent a priority in the current phase of the research process.
On the basis of the findings obtained in this study, it is argued that, by relying on a
range of pre-established demographic characteristics and rules, the proposed user
model tool could enhance the precision of advertisement targeting and therefore boost
business sales and profitability. Adaptive advertising systems could be made more portable by the proposed user model, which is compatible with any website. Thus, hypotheses H9 and H10 regarding the ease of incorporation of the user model tool in any JSP
website hold true. Furthermore, the tool could not only help existing systems to expand,
but also supply the flexibility needed by businesses to achieve advertisement customisation whilst eliminating the necessity of current model revamping. Additionally, the
model could allow websites to request a method that can be found in a particular location on those websites, whilst at the same time organising and amending all the advertisements on a specific page with the help of a single code and in conformance with
established adaptation rules [36] as well as maintaining a record of the advertisements
that have been shown and accessed by users. To some degree, these aspects are in
agreement with, and therefore validate, hypothesis H11, which maintained that stereotypes could be easily comprehended, applied and updated by all administrators of websites.
6
Discussion
As stated in the above hypotheses, the of the user model structure introduced is considered suitable for advertisements by the participants to the study, who agree that it can
be constructed and used in an easy way. The most popular features were the portability,
and thus the lightweight structure of the data representation, in terms of export facility
in XML, and the multi-system login facility, as well as connection with social networks,
via the introduction of the social networks login. Importantly, users considered that the
user characteristics that are collected by the proposed user model positively affect the
selection of the appropriate advertisements for them. Moreover, the additional user
model information that is obtained by tracking the users positively affects and enhances
the delivery of appropriate advertisements.
Courses, multimedia, and advertising, among others, can all be adapted for users by
many adaptation systems designed for this purpose. Indeed, ADE, ADROSA, ADWIZ,
MyAds, and more recent MyAds are all examples of adaptive systems. ADWIZ [13]
possesses a trivial user model structure. It relies only on search keywords supplied by
a user to a search engine. Similarly, MyAds [14] possesses a trivial user model structure
since a receptor component receive the user profiles (no storage for data) through several means, including wireless, and evaluates which content should actually be allocated to which displays. The AdROSA system depends on the categorisation of web
banners for groups, based on similarities between individuals, it only processes a minimum of data about users, from the local session, to categorise them, so as such, it is
also a relatively trivial user model. Such trivial implementations for the user model
don’t allow for enough information to make useful decisions about recommending the
appropriate advertisements to users. The user model in these three systems represents
short-term interests, since it uses fixed static characteristics of the users, like search
words. The design of the user model in AEADS is attained by separating it into four
components, in order to make the adaptation process easier and reusable. The user’s
data is arranged as three components: user’s data, behaviour data, and social data. This
type of construction and the permanency of data support allows the system to predict
not only short-term, but also long-term, for future sessions, the desired advertisements
that will be relevant to users, based on businesses rules. The future advertisement is the
fourth component that contains advertisements that will be shown in the future, at the
next login, to each user.
Another example of an adaptive system, ADE, as a generic adaptive delivery system,
supports a rich user model, which runs however in-session only. It often is used to deal
with user model parameters such as: how many times a user has visited a concept; the
active time spent by the user on each page; and clicks on links, among others. However,
the ADE user model is non-persistent between sessions, which is something that was
considered essential for the business case. Moreover, ADE runs as a standalone system,
which was not considered appropriate for the paradigm used in this research, which
needed to allow businesses to use their original websites, enhanced with AEADS features. On the other hand, the user model in the more recently developed, standalone
system MyAds [37] contains information concerning the buyers and their viewing history - the ads they have seen. It is initialised by registration process or Facebook login,
before being updated by the user’s clicking action, to assign a related advertisement to
the current user. The users are then able to declare their specific interests, by labelling
categories with numbers from 1 to 10. Later on, they can annotate the recommendations, to specify if they would be willing to purchase such items or not.
Overall, in the user model in AEADS all functions of the user model are separated
from storage and located within the delivery section of the AEADS system. This separation aims to increase portability and should allow an easy extension of the system,
without affecting the system as a whole. AEADS can additionally deal with users that
have no identification, as well as with users with identification. AEADS uses two algorithms for each of the two types and, in the case of those users that have no identification, the system starts by showing all advertisements randomly, and then monitoring
the user’s clicks on each advertisement, in order to allow some adaptation. The comparison of the different adaptive systems discussed in this section is summarise in the
table below.
Table 3. User Model in Different Systems.
Purpose
UM Size
UM Initialisation
UM
Structure
UM
Maintenance
UM Implementation
Advertisement
ShortTerm
LongTerm
Social
Login –
Registration - Automatically
4 components represent levels of information
Modifier
Engines
Separated
in Delivery model
Inference
Engines
Separated
in Delivery model
ADE
Courses
ShortTerm
Registration
Any user
data (variablevalue)
Presentation handler
Adaptation engine
AdROSA
Advertisement
ShortTerm
Automatically
vectors
database
Web (ad)
usage
mining
Web (ad)
content
mining
ShortTerm
Get only
recent activity –
search
word
Trivialsearch
word
---
Ad server
System
AEADS
ADWIZ
MyAds
recent
MyAds
Advertisement
Advertisement
ShortTerm
Automatically
Trivial receptors
----
Advertisement
ShortTerm
LongTerm
Social
Login –
Registration -
2 components represent users and
companies
----
Information
Processing
Component
Ads Contents
Manager
on server
side
7
Conclusions and Future Work
In this paper, a lightweight user modelling approach has been proposed. It could help
Internet users to register to any web-based e-commerce system, and thus help companies’ access their target audience more directly, by tailoring their marketing campaigns
towards specific consumer demographics and focusing their advertisements on users
who satisfy a predetermined range of criteria. Based on the outcome of theoretical considerations and practical testing, a minimum set of user model dimensions have been
validated. The evaluation results indicate that the initial functionality and usability of
the small prototype system is promising. Further modifications are planned, based on
the suggestions offered by survey respondents. The user modelling tool can be refined
further, by taking into account user feedback and creating a lightweight adaptive system
that is more customisable, and based on the needs and preferences of Internet users. As
an immediate next step, in our follow-up studies, the delivery engine has been implemented, which is resident on the same website server, to deliver the advertisements to
Internet users. This part parses the contents in the XML file and uses adaptation strategies to send the appropriate advertisements to the appropriate user, based on their user
model.
References
1. InternetAdvertisingBureau: Internet Advertising Revenues Report 2012 full year
results. (2012)
2. Halasz, F., Schwartz, M., Grønbæk, K., Trigg, R.H.: The Dexter hypertext reference
model. Communications of the ACM 37, 30-39 (1994)
3. De Bra, P., Houben, G.-J., Wu, H.: AHAM: a Dexter-based reference model for
adaptive hypermedia. Proceedings of the tenth ACM Conference on Hypertext and
hypermedia: returning to our diverse roots: returning to our diverse roots, 147-156.
ACM (1999)
4. Cristea, A.I., de Mooij, A.: LAOS: Layered WWW AHS authoring model and their
corresponding algebraic operators. WWW03 (The Twelfth International World
Wide Web Conference), Alternate Track on Education, Budapest, Hungary, (2003)
5. Brusilovsky, P.: Adaptive hypermedia: An attempt to analyze and generalize. In:
Brusilovsky, P., Kommers, P., Streitz, N. (eds.) Multimedia, Hypermedia, and
Virtual Reality Models, Systems, and Applications, vol. 1077, 288-304. Springer
Berlin Heidelberg (1996)
6. Kobsa, A., Koenemann, J., Pohl, W.: Personalised hypermedia presentation
techniques for improving online customer relationships. The knowledge
engineering review 16, 111-155 (2001)
7. Brusilovsky, P., Millán, E.: User models for adaptive hypermedia and adaptive
educational systems. The adaptive web, 3-53. Springer-Verlag (2007)
8. Bra, P.D., Calvi, L.: AHA! An open Adaptive Hypermedia Architecture. New
Review of Hypermedia and Multimedia 4, 115-139 (1998)
9. Stash, N., De Bra, P., Cristea, A.: Adaptation to learning styles in a general-purpose
system AHA!(Adaptive Hypermedia Architecture). Educational Technology &
Society 11, (2008)
10. Smits, D., De Bra, P.: GALE: a highly extensible adaptive hypermedia engine.
Proceedings of ACM-CHH 63-72 (2011)
11. Scotton, J., Stewart, C., Cristea, A.I.: ADE: The Adaptive Display Environment for
Adaptive Hypermedia. Proceedings of the ACM Hypertext 2011 International
Conference, Eindhoven, The Netherlands, (2011)
12. Kazienko, P., Adamski, M.: AdROSA—Adaptive personalization of web
advertising. Information Sciences 177, 2269-2295 (2007)
13. Langheinrich, M., Nakamura, A., Abe, N., Kamba, T., Koseki, Y.: Unintrusive
customization techniques for Web advertising. Computer Networks 31, 1259-1272
(1999)
14. Di Ferdinando, A., Rosi, A., Lent, R., Manzalini, A., Zambonelli, F.: MyAds: A
system for adaptive pervasive advertisements. Pervasive and Mobile Computing 5,
385-401 (2009)
15. Brailsford, T.J., Choo, B.S., Davies, P.M.C., Moore, A., Stewart, C.D., Zakaria,
M.R.: Web-based Hierarchical Universal Reactive Learning Environment
(WHURLE): An Overview (White Paper for the WHURLE Project). (2001)
16. Mérida, D., Fabregat, R., Marzo, J.-L.: SHAAD: Adaptable, adaptive and dynamic
hypermedia system for content delivery. Workshop on Adaptive Systems for Web
Based Education. WASWE2002. Málaga España. Citeseer (2002)
17. www.google.com/adsense/
18. Davis, H.: Google advertising tools: Cashing in with AdSense, AdWords, and the
Google APIs. " O'Reilly Media, Inc." (2006)
19. http://w3techs.com/technologies/overview/advertising/all
20. Kobsa, A.: Generic user modeling systems. The adaptive web, 136-154. Springer
(2007)
21. Al Qudah, D., Cristea, A.I., Shi, L., Al-Sayyed, R.M., Obeidah, A.: MyAds: a social
adaptive system for online advertisement from hypotheses to implementation.
(2014)
22. Faust, K.: Very local structure in social networks. Sociological Methodology 37,
209-256 (2007)
23. TheStatisticsPortal: Number of monthly active Facebook users worldwide 20082014. (2015)
24. Hof, R.D.: Facebook's new ad model: You. FORBES 188, 106-+ (2011)
25. Cristea, A.I.: What can the semantic web do for adaptive educational hypermedia?
Educational Technology and Society 7, 40-58 (2004)
26. Wu, H.: A reference architecture for Adaptive Hypermedia Applications.
Technische Universiteit Eindhoven (2002)
27. http://www.rci.rutgers.edu/~cfs/472_html/Planning/PlanRecog.html
28. Benaki, E., Karkaletsis, V.A., Spyropoulos, C.D.: Integrating user modeling into
information extraction: The UMIE prototype. User Modeling, vol. 97, 55-57.
Springer (1997)
29. Qaffas, A., Cristea, A.: How to create an E-Advertising Domain Model: the AEADS
approach. The 2014 International Conference on e-Learning, e-Business, Enterprise
Information Systems, and e-Government(EEE'14), Las Vegas, United State (2014)
30. Qaffas, A., Cristea, A.: How to Create an E-Advertising Adaptation Strategy: The
AEADS Approach. In: Hepp, M., Hoffner, Y. (eds.) E-Commerce and Web
Technologies, vol. 188, pp. 171-178. Springer International Publishing (2014)
31. Benaki, E., Karkaletsis, A., Spyropoulos, D.: User modeling in WWW: the UMIE
Prototype. Proc. of the Workshop” Adaptive Systems and User Modeling on the
World Wide Web, 6th International Conference on User Modelling UM, vol. 97,
(1997)
32. McIver, J., Carmines, E.G.: Unidimensional scaling. Sage (1981)
33. Gliem, J.A., Gliem, R.R.: Calculating, interpreting, and reporting Cronbach’s alpha
reliability coefficient for Likert-type scales. Midwest Research-to-Practice
Conference in Adult, Continuing, and Community Education (2003)
34. Kazienko, P., Adamski, M.: Personalized web advertising method. Adaptive
Hypermedia and Adaptive Web-Based Systems, 146-155. Springer (2004)
35. De Bra, P., Smits, D., Van Der Sluijs, K., Cristea, A., Hendrix, M.: GRAPPLE:
Personalization and adaptation in learning management systems. World Conference
on Educational Multimedia, Hypermedia and Telecommunications, vol. 2010,
3029-3038 (2010)
36. Stash, N., Cristea, A.I., De Bra, P.: Adaptation languages as vehicles of explicit
intelligence in Adaptive Hypermedia. International Journal of Continuing
Engineering Education and Life Long Learning 17, 319-336 (2007)
37. Al Qudah, D.A., Cristea, A.I., Hadzidedic, S., AL-Saqqa, S., Rodan, A., Yang, W.:
Personalized E-advertisement Experience: Recommending User Targeted Ads.
12th International Conference on e-Business Engineering, Bejin, China (2015)