An Expandable Recommendation System on IPTV

advertisement
An Expandable Recommendation System on IPTV
Jie Xiao and Liang He*
Instituation of Computer Aplication, East China Normal University, Dongchuan Road.500,
200241, Shanghai, China
{ jxiao@ica.stc.sh.cn, lhe@ica.stc.sh.cn }
Abstract. IPTV’s mass resources, high quality service and its open and free
interactive service mode attract billons of user. It has brilliant potential and also
has great challenge. Traditional one algorithm supported recommendation
system is no longer content IPTV’s huge service demand. In order to solve this
problem, an expandable recommendation system is proposed in this paper. It
contains new flexible system framework and basic ideas of applicable
algorithms. And it has been implemented in real IPTV system and works well.
Keywords: IPTV, Recommendation Framework, Recommendation Algorithm,
tag, template.
1
Introduction
IPTV (Internet Protocol Television) [1] is the television broadcast service based on
Internet Protocol. IPTV take advantage of digital, computer, internet and broadcasting
technology and combines them into one platform. It provides digital broadcast, video,
information, entertainment and e-commerce services though broadband on television
or computer. Billions of users are attracted by IPTV’s mass resources, high quality
service and its open and free interactive service mode. It has becomes a most
promising global business.
The biggest characteristic of IPTV is interactive service. So providing good
personalized service [2] for IPTV user is significant to its development. At present,
most of the IPTV system is using the recommendation system which has been long
time used in the e-commerce system. And most of these systems only use one
algorithm in the whole recommendation process. These algorithms have lots of
advantage. But for IPTV’s multiple interactive scenarios, is only one algorithm
enough to content all demand? The answer is obvious. In order to settle multiple
service requirements, in this paper, a flexible and expandable recommendation system
is proposed. It can adopt different algorithms for different user scenario. And all the
algorithms and scenario is expandable. Collaborative filtering is widely used in
recommendation system. However, it has limitation in selecting the most suitable
nearest neighbor. So in this recommendation system, tag-based collaborative filtering
is proposed. Besides, we proposed user template algorithm to handle new user and
2
Jie Xiao and Liang He*
overload problem. And we also provide some small current video based
recommendation algorithms.
2
Related Works
Personalized recommendation is one important part in user behavior analysis. In
simple terms, it’s a process of searching the resources which the user might interest in.
This technology has developed for decades. At present, Collaborative Filtering,
Content-based Recommendation, and Association Rule-based Recommendation are
the most matured and popular recommendation algorithms.
Collaborative Filtering [3] is the hottest and successful technology implemented in
recommendation system. The user-based collaborative filtering is one most used
algorithm. It assumes if user’s rating is similar in lots of items, their rating on other
items will be similar, too. So the user’s rating for the items is predicted by neighbors’
weighted average values [4].
Content-based Recommendation [5] assumes if user is interest in an item before,
user will interest in the similar items, too. It usually using vector space model [6] to
describe user’s interest.
The aim of Association Rule-based Recommendation [7] is to find the relations
between items, such as the famous story about beer and diaper. The rules are set by
people or mining technology.
Table 1. Comparison of three algorithms
Recommendation
Technology
Advantage
Disadvantage
Collaborative
Filtering
Easy to find new interest;
Able to identify the quality
of resource; Able to use on
multi-media resources,
No cold-start and sparsity
problem; Easy to find new
or uncommon resource.
Cold-start
problem; System owns certain
Sparsity problem; Slow amount of user rating
in overload system.
but not overload.
Content-based
Recommendation
Cannot find new interest
or identify the quality of
resource;
New
user
problem; Only suit for
text resource.
Association Rule- Simple; Able to find new Hard to set the rules; Not
based
interest.
suitable
for
content
Recommendation
changeable system.
Applicable system
Text system which
easy
to
extract
characteristic of items
and
owns
certain
amount of user record.
System which content
is not change much.
Table1 shows that these three algorithms have their different advantage,
disadvantage and applicable scope. Only choose the right algorithm in different user
scenario can obtain the best recommendation result. So that is the aim we propose an
expandable recommendation system.
An Expandable Recommendation System on IPTV
3
3
System Framework
IPTV provide multiple interactive services. When user just login, the user might to
select one video he or she interested to watch. In this scenario, system needs to
recommend according to user’s overall interest. And when user is watching a video,
he or she might want to watch related videos. Or when user is shopping or playing
online game, the user need other personalized service. In order to content IPTV’s
multiple service demand, we need a flexible recommendation framework which able
to: 1) Load and run kinds of algorithms; 2) Select different algorithms according to
different user scenario; 3) Configurate the algorithm’s weight and its arguments. 4)
Handle the unstructured source data into structured data.
Fig. 1. IPTV Recommendation System
Fig.1. shows all components and overall flow of the system. It divided into the
following steps:
1. Source data processing.
Source data includes static and dynamic data. Static data contains user register data
(user name, gender, age, hobby, and etc.) and history data (user watching record and
system record). Dynamic data is user’s current behavior (current user scenario,
play/pause the video, and etc.). The system processes these unstructured data into
structured data which can be handled by the other component of the system.
2. Algorithm calculating and mixing.
This step divided into offline and online part. Some of algorithms contain both part
and others only contain one part. The offline component handles the offline
algorithms. The online component handles the online algorithms. And they both have
4
Jie Xiao and Liang He*
algorithm mixer which is the most important component in the system. Below is the
structure of it.
Fig. 2. Algorithm Mixer
Scenario means user’s different behavior. For example, “just login”, “watching
video”, “shopping” is one kind of scenarios. There are listed in the scenario store area.
Each scenario’s algorithm configuration is stored in the configuration store area. One
scenario can load one or multiple algorithms. When one scenario is selected, the
mixer will run the according settings.
Weight and amount is two mutually exclusive fields. They represent different mix
mode. In weight mode, each algorithm can be set in high, mid or low weight.
Recommendation item from different algorithm will be given different score rank.
And the mixer selects the top-n score items as the last result. If item’s score from one
algorithm is much higher than others, most of the items get from other algorithm
cannot be displayed to user. But sometimes, we hope user get comprehensive result.
Amount mode solved this problem. In this mode, the mixer will generate the result by
the set amount.
3. Recommendation generation.
After mix, the system will collect all the recommendation result and convert them
into visual message and output to the screen.
Administrator control platform provide interface to set kinds of system arguments
including loaded algorithms, loaded scenarios and detail arguments of algorithm.
An Expandable Recommendation System on IPTV
4
The Basic Idea of Recommendation Algorithms
4.1
Tag-based Collaborative filtering
5
Traditional Collaborative filtering recommend through nearest neighbor. But it has
two problems:
1)
It couldn’t really understand the preference of users. Take movie
recommendation as example. User gives high score to “Avatar” means user
like it. What attract the user? Its 3D effects or its science fiction plot? We
don’t know the reason. The preference of user is vague. [8]
2)
It doesn’t find the real nearest neighbor. For example, boy and girl both are
programmers and both like to watch high-tech news. They become nearest
neighbor according to the algorithm. But besides high-tech news, boy likes
to watch NBA games and girl likes to watch soap opera. Recommend NBA
game to girl is obvious not appropriate.
In order to overcome these two drawbacks, Tag-based Collaborative filtering is
proposed in this paper. New “user-catalog-tag-resource” user interest model is
adapted on the base of traditional “user-resource” model.
Fig. 3. User interest model
User interest is described by five catalogs. Each catalog contains lots of tags and
has different weight (a%-e%). Each tag has scores. All the catalogs and tags can be
expanded as we want.
First step is to find the nearest neighbor in each catalog. The following formula is
to calculate user ui and user uj’s similarity.
sim(ui , u j , ck ) 
Cik  C jk
Cik  C jk
Cik is the tag set ui watched in catalog Ck.
(1)
6
Jie Xiao and Liang He*
The following step is similar to the traditional collaborative filtering. After get the
top-k nearest neighbors, predict the user’s rating by neighbors’ weighted average
values. And then recommend the top-n score resources.
4.2
User Template
As we know, collaborative filtering cannot get good recommendation result when the
record of user or system is rare. And when the system is overloaded, the calculation
makes the system rather slow. In order to handle these two problems, user template
recommendation is proposed.
As Fig.4 shows, it classify user into different type by user’s characteristic and
provided different recommendation service for them. In simple terms, user template
can be described as:
user with characteristic A, B, …, K
might like
resource with characteristic A, B, …, N.
There are two methods to build user template. One is manual modeling by expert’s
experience and statistic of video service system. It has wide universality and
authority. But internet is changeable. Manual modeling is not enough. So we
proposed automatic modeling. It’s an algorithm similar to decision tree [9]. For each
user classification, it makes statistic analysis to the characteristic of videos which user
has watched. Then it selects the top-n video characteristic as candidate template.
Finally, the template which score are higher than the set threshold is selected as user
template.
Fig. 4. User classification
4.3
Resource-based Recommendation
Besides the above algorithm, this paper also provides some simple algorithms using
for the “watching video” scenario. They make recommendation based on the current
video.
─ Video similarity recommendation. It will recommend the video similar to
current video. For example, same actors, director, style or keyword.
An Expandable Recommendation System on IPTV
7
─ Current neighbor recommendation. Current neighbor means other users who
watched the current video and rate it with high scores. It recommends the
current neighbor’s next watching video.
─ Catalog hot recommendation. It will recommend the hottest video which in the
same catalog of the current video.
5
System Implementation
All above framework and algorithm proposed is from our institution’s science project
cooperating with one of the biggest IPTV providers in Shanghai which own millions
of users and video resources. And this expandable recommendation system has been
implemented in their IPTV system. The system language is Chinese. Fig.5 is one
system screenshot. The red highlight part is the recommendation area. It performs
well in the real applications both in recommendation result and system efficiency.
Fig. 5. System screenshot
6
Future Work
In this paper, the author creates an expandable recommendation framework and
demonstrates the basic ideas of some algorithms. The implement system works well.
But there is also lots of work need to be done in the future.
First, the paper is emphasis on the recommendation framework and just gives
general ideas to algorithms. These algorithms still have lots of potential need depth
research. And there must be more algorithms suit for IPTV user scenarios. We have to
devote our effort to discover.
Second, the framework is ready. How to choose the right algorithms and
configuration to achieve the best recommendation performance is need us to explore.
8
Jie Xiao and Liang He*
Acknowledgement
This work is supported by a grant from the National High Technology Research and
Development Program of China (863 Program) (No.2011AA01A107) and the
Shanghai Science and Technology commission Foundation (No.10dz1500103).
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
Yang Xiao, Xiaojiang Du, Jingyuan Zhang, Fei Hu, Guizani: Internet Protocol Television
(IPTV): The Killer Application for the Next-Generation Internet. In: Communications
Magazine, 2006, vol 45, pp. 126-134.(2007)
Chun Zhen, Chunxiao Xing, Lizhu Zhou: Overview of Personalized Service Techniques.
In: Journal of Software, 2002, vol.13, pp.1953-1961(2002).
J.B. Schafer, D.Frankowski, J.Herlocker, S,Sen: Collaborative filtering recommender
systems. In The Adaptive Web: methods and strategies of web personalization, pp.291324(2007)
IE Shparlinski: On some weighted Average Values of L-function. In: Bulletin of the
Australian Mathematical Society, 79 : pp.183-186(2009)
Michael J. Pazzani, Daniel Billsus: Content-Based Recommendation Systems. In: The
Adaptive Web: Lecture Notes in Computer Science, pp.325-3441(2007)
G. Salton, A. Wong, C. S. Yang: A vector space model for automatic indexing. In:
Communications of the ACM, v.18, pp.613-620(1975)
J. J. Sandvig, Bamshad Mobasher, Robin Burke: Robustness of collaborative
recommendation based on association rule mining. In: Proceedings of the 2007 ACM
conference on Recommender systems, pp.105-112(2007)
Jie Xiao, Liang He: Keyword weight adjusting schema based on domain repository. In:
3rd IEEE International Conference on Computer Science and Information Technology, pp
221-225(2010)
Banfield R.E., Hall L.O., Bowyer K.W., Kegelmeyer W.P.: A Comparison of Decision
Tree Ensemble Creation Techniques. In: Pattern Analysis and Machine Intelligence, IEEE,
vol.29, pp.173-180(2007)
Download