You are on page 1of 5

JOURNAL OF COMPUTING, VOLUME 3, ISSUE 6, JUNE 2011, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 28

An Efficient Data Model and Vector Space
Model Based Dynamic TV-recommender
system for Interactive television
Md. Ariful Islam Khandaker, Abdul Kadar Muhammad Masum, Golam Mostafa
AbstractHuman interest learning is a complicated process as it changes in different situation and factors. Hence designing
efficient algorithm for the retrieving of interesting television events for user by leaning the intention or interest of the TV-owner
dynamically is a big challenge. This paper presents a novel approach of TV-recommender system using Vector Space Model.
Effective data model have been proposed to successfully storage of TV events information which is vital for developing the
recommender system. The experiments show that the algorithm has the ability to retrieve effective TV program according to the
user interest. We consider the learning rate to tune the system as short term learning and long term learning which proves the
user more control over the television to filter most important TV-events.
Index TermsTV-recommender system, iTV, EPG, Vector Space Model, Short time learning, Long time learning.



1 INTRODUCTION
NTERACTIVE television (iTV) is a modern vision of
ubiquitous world. The television is more dynamic with
many more functionalities. Besides other features, it is
intelligent enough to learn the TV users interest level.
The interactive TV can learn the viewer viewing pattern.
And it stores the information (user profile) that the user is
more interested to view. On the basis of interest it can
suggest TV events for the user. Figure A shows the con-
ceptual model of an interactive television. The TV set in-
cludes a Personal Video Recorder (PVR) or Set Top Box
(STB) which have the capabilities to store some TV-
events. Intelligent software is integrated with the
PVR/STB which can collect either the information of TV-
events from the service provider by observing user re-
sponse to the TV set. Whenever a user selects a TV pro-
gram, the system immediately collects the Electronic Pro-
gram Guide (EPG) data of that TV-event from different
source of WWW and performs some operations to build
the user profile. As long as the user changes different TV-
events, the system updates the user profile. The recom-
mender system finds the interesting TV-events on request
of user command on the basis of stored user profile.







Fig. 1. A conceptual model for interactive television that
have the recommend system.

2 LITERATURE REVIEW
Some work is being carried out in the broader area of re-
commender systems in general. Their purpose may be to
suggest e.g. web pages [7], [8] or music hands [9]. Besides
the goal of alleviating human-computer interaction, re-
search in automated user profiling is also driven by the
commercial product model. Focusing on TV recommend-
ers, there exist several proposals that examine implicit
profiling: electronic program guides [6], PTV [6] and a
multi-agent TV recommender [7]. In most of the current
commercial PVR, users have to specify their interests by
themselves to get high quality recommendations. The
drawback is that user intervention is required, and that
the profile is henceforth static. In contrast, PTV [10], the
multi-agent TV recommender [3], and the electronic
I

- Md.Ariful Islam Khandaker is with the Department of CSE , International
Islamic University Chittagong,Bangladesh-4203.
- Abdul Kadar Muhammad Masum is with the Department of DBA, Inter-
national Islamic University Chittagong, Bangladesh-4203.
- Golam Mostafa is with the Department of CSE, Stampford University,
Bangladesh.



User
PVR/STB Service
provid-
ers
Intelligent program
to interact with
TV set
With rec-
ommend
software
EPG
www
User response/evaluation
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 6, JUNE 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 29

guide described in [6] enable automatic design of the pro-
file. ITV [10] is a client-server system operating over the
web. It uses an explicit profile and a collaborative algo-
rithm [11] to make recommendation. Thus the system
have to face all the problems of collaborative algorithms
[12], such as new user, sparse rating, scaling problem. The
multi-agent TV recommender [3] and the electronic pro-
gram guide [6] implement both an explicit and an implicit
profile using content-based filtering. The implicit profile
is made automatically, not requiring any initialization
from the user. The implicit profile of the multi-agent re-
commender [7] is built from TV programs. Positive ex-
amples are created starting from wholly viewed pro-
grams, and for each of them, a negative example is
created. Profile learning and program classification may
be carried out using a Bayesian classifier or a decision
tree, such as proposed in [7]. The authors conclude on the
similarity of experimental results from these two tech-
niques. The electronic guide [5] also uses the metadata of
TV programs to construct implicit profiles. Clustering of
examples relies on an ontology which extends the DVB-SI
standard with super and sub-categories: the profile is
then represented by the class containing positive exam-
ples.
3 REPRESENTING TV-EVENTS
TV-event representation is very is another challenge be-
cause if we can represent the TV-event in proper way
then we can able to design recommendation algorithm.
The tv-anytime[1] forum is the research group responsi-
ble for designing multimedia information like tv program
in convenient way for global access. By following their
strategy we have proposed our data model for the TV-
recommendation system. We have studied tv guide from
different newspaper and also from online and observed
that they have some structure to describe the tv show.
Figure 2 depicts the overall structure of typical TV-events.

-According to the following observation we can easily
describe the TV-events as some common attributes. We
can represent tv-programs in xml as set of vectors with
attributes. Here is the vector representation for above
three TV progrms respectively.

P1=G: movie, S; comedy, P: Chirs Columbus, P: Robin
Williams, P: Sally Field, K: Academy award.

P2={G:Sport,S:Tennis,P:Novak Djokovic,P:David Fer-
rer,K:Grand Slam}

P3={G:Music,S:Jazz,P: Phill Collins,P: Black Eyed Peas,P:
Chic,P: Nile Rodgers,K:Montreux, K:Jazz Restival }

Where
G stands for Genri
S stands for Sub-genri
P stands for tv-person
K stands for keywork



: is used to separate the attribute and the type of
attribute f you are using Word, use either the Microsoft
Equation Editor or the MathType add-on
(http://www.mathtype.com) for equations in your paper
(Insert | Object | Create New | Microsoft Equation or
MathType Equation). Float over text should not be se-
lected.
4 VSM BASED ALGORITHM
According to the specification of TV-anytime forum and
some analysis we have found that a TV-event data can be
represented as a set of terms that falls in any item from the
following set.
S= {Genre, Sub-genri, TV-person, Keyword},
We can set weight to each term according to the rule
WGenri>WSub-genre>WTV-person>WKeyword.
The algorithm has two phases. First one is learning phase
and second one is similarity measurement phase.
4.1 Learning Phase
For each TV-event viewed by the viewer the system will
consider the features of that TV-event.
(Case-1) If the TV event term is already in the user
profile database then the weight of the features will be
updated in proportion to the feedback


(1)
Where
i
w


i
w'
is the weight of term ti after update, and
wi is the weight of term ti before update.
( ) 1 0 s s o o
is the learning rate that determines
how quickly the user profile forgets old preferences and
tracks new ones,
|
is the ratio of users real watching time (Tr) to the
Main-category/genri
(Example: Movie, Sports, Music, Talk Show

Sub-category/ sub genri
(Action, Horror as types of movie)
(Football, Tanis as types of sports) and so on
List of person like actors, director, name of players or
name of singers etc
Some keyword that best represent the program like his-
torical place and name of championship for sports
Fig. 2. Generalized structure for tv-show
description described by tv-anytime
) ( w ) 1 (
i
i f w
i
+ = ' | o o
| | 1 , 0 e =
t
r
T
T
|
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 6, JUNE 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 30

0 20 40 60 80 100 120
0.1
0.2
0.3
0.4
0.5
Observation of the weights in a genri:Music
Number of sample value
W
e
i
g
h
t
programs total duration time (Tt).
|
can be considered
as the users evaluation to the program which he/she has
viewed.

(Case-2) if a term does not exist in the profile,
Calculate the terms weight
) (i f w
i
= | o (2)
Here Wi is the weight of the new term ti and
|
have the
same meanings as case1.

4.2 Similarity Measurement Phase
We adopt the Vector Space Model (VSM)[13] as the fea-
ture extraction and object information presentation me-
thod. Each term is defined as a 2-tuple (term, weight). So
the user profile can be represented as a vector of these 2-
tuples, if there are m distinct terms in the profile, then it
will be represented as a vector:
)) , ( ),......., , ( ), , ((
2 2 1 1 m m
w t w t w t P =

) 1 ( 1 m i w w
i i
s s + >
(3)
where ti is a term, wi is the weight of term i . So the us-
ers profile can be conceptually represented as the follow-
ing vector:
P = (w1,..., wn) (4)
where wi is the weight of term ti in the profile. Similar-
ly, a program can be also represented as a vector with n
items, which are the same as those in the profile vector,
that is term ti in the profile vector and content vector is
the same:
C = (u1,..., un

) (5)
Where ui is the weight assigned to term ti .
A commonly used similarity metric is the cosine of the
angle between the two vectors. Given a program C
=(u1,...,un) and a profile P =(w1,...,wn) , the cosine simi-
larity can be calculated as follows:
Sim(C,P) =
|| || || || P C
P C

=
=
=
=
=
=

n i
i
n i
i i
i i
n i
i
i i
w u
w u
1
2 2
1
(6)


5 RESULT AND EXPERIMENTAL ANALYSIS

For evaluating the performance of our system we demon-
strate several experiments. We have written code in C to
generate random data to simulate the different users.
Separate programs are used to develop user profiling and
measure similarity to find useful TV-events. The experi-
ments are described bellow.

5.1 Experiment-1: Does the system work?
For the verification of our algorithm we have used 10
standard programs. We demonstrate a program written
in C++ to make the user profile as combination of terms
and weight. We took the profile with certain threshold
value. With another program we made the similarity
measurement with some other three new TV-events. We
found that the algorithm can built the user profile (list of
attributes and weights) which reflects the user interest
from initial set of TV-program data. Moreover on the ba-
sis of user profile the system can find similar programs
hierarchical order from new TV-program set.

5.2 Experiment-2: How about the quality of learning
process?
We have run our user profiling with 500 EPG. And for the
attribute Music as genri we have observed the weights as
following figure.






Fig. C. Observation of the weight updates in a genre: music
The plotted figures represents the weight values of music
programs of a particular user interacting with TV, that
represents the interest level. The main observations of the
experiments-2 are:
1. The system can learn quickly as it takes only 12 to 15
music type tv-program to learn the level of interest
for the music. Because the label of interest of music is
0.3 to 0.5 and the system can reach this range after 12
to 15 tv-program data related to music.
2. By the change of user interest for particular type of
TV-program, the weight values of attribute in the
user profile concerning the tv-program also change.
But the weight does fluctuate out of range. This
indicates that the system is stable and takes time to
judge the changes in interest label.
3. For finding the similarity the system takes the current
value of weights for each attributes. As the weight
values always remain in certain range so the
probability to find similar item is close to the user
profile values and very near to accurate.

5.3 Experiment-3: The effect of the learning rate
value
We also try to observe how the learning rate (
o
) act for
changing weights in user profile. For this analysis we
have used two different experiments with different
number of dataset

Subexperiment 1
Genri Music
Number of Sample
EPG
190
Learning Rate( o ) 0.1, 0.2 and 0.3

JOURNAL OF COMPUTING, VOLUME 3, ISSUE 6, JUNE 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 31

0 20 40 60 80 100 120 140 160 180 200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Observation of the weights in a genri:Music
Number of sample value
W
e
ig
h
t


o= 0.1
o= 0.2
o= 0.3
0 20 40 60 80 100 120
0
0.2
0.4
0.6
0.8
1
Observation of the weights in a genri:Music
Number of sample value
W
e
i
g
h
t


o= 0.3
o= 0.4
o= 0.5
Subexperiment- 2
Genri

Music
Number of sample
EPG
110
Learning Rates( o ) 0.3, 0.4 and 0.5
Fig D: Parameters of experiments 3
Figure E1 and Figure E2 show the effect of learning rate
for the change of a genri Music in user profile for two
different set of data.






Fig. E1. Effect of learning rate for changing the weights in terms of
sample data case study 190 EPG data.(Long Time Learning).







Fig. E2. Effect of learning rate for changing the weights in terms of
sample data. Case study 110 EPG data.(Short Time Learning).
The main observations of experiment-3 are
1. The value of learning rate is much effective when we
want the system to learn for long time with large da-
taset. That means when the system continues the
learning process for a particular user for long time
(long time learning) then the learning rate should be
kept less for meaningful and efficient output for the
recommendation system.
2. When the system learns the user interest for short
time (short time learning) then the value of learning
rate does not affect much for finding the user interest
and measure the similarity
6 CONCLUSION
The algorithm is able to learn the user interest quick-
ly. For each vectors of program data it stores the
weight and after executing several (12 to 15) data it
reaches the right range of value for that vector. If the
change of interest it also takes same number of pro-
gram data to change the interest level. The system
does not fall immediately. As for similarity measure-
ment the system needs the current profile data, and
the weight changes within some range which
represents the interest of some TV-user, so most of the
time it gives the correct output program list from in-
coming program. For up to 120 tv-programs the
choosing the learning rate is not too much sensitive.
But when the system learns the user interest for long
time more than 200tv-programs then the choosing
learning rate is very important. On that case learning
rate should be kept smaller for the better system per-
formance. The main advantage of our recommender
system is that it is very simple and hence takes very
less time of complexity as we work only with
attribute and weigh for computation. Moreover it re-
quires very less memory space and also it is secured
system as everything is working on the users personal
TV.
REFERENCES
[1] TV-Anytime Environment Requirements Document, TV035r6, TV-
Anytime Forum, Aug. 2000, http://www.tv-anytime.org/
[2] TV-Anytime System Requirements Document, TV036r2, TV-Anytime
Forum, Apr. 2000.
[3] K. Kurapati, S. Gutta, D. Schaffer, J. Martino and J. Zimmerman. A
Multi-Agent TV Recommender. In Proceedings of the User Model-
ing 2001: Personalization in Future TV Workshop, Sonthofen, Ger-
many, 2001.
[4] Joo Ascenso, and Alberto Silva. Software Agents for Digital Televi-
sion. In Proceedings of International Conference
[5] DVB. Digital video broadcasting. http://www.dvb.org,2000
[6] L. Ardissono, F. Portis and P. Torasso, F. Bellifemine, A. Chiarotto and
A. Difino, Architecture of a system for the generation of personalized
Electronic Program Guides.
[7] H. Lieberman, Letizia: An Agent That Assists Web Browsing, in
Chris S. Mellish, editor, proceedings of the Fourteenth International
Joint Conference on Artificial Intelligence (IJCAI-95), pp. 924-929,
Mon treal, Canada, 1995, Morgan Kaufmann publishers Inc,: San Ma-
teo, CA, USA
[8] M. J. Pazzani, J. Muramatsu and D. Billsus, "Syskill Webert: Identify-
ing Interesting Web Sites", inAAAI/lAAl, Vol. 1,pp. 54-61, 1996.
[9] U. Shardanand and P. Maes, "Social Information Filtering: Algorithm
for Automating "Word of MauW", in Proceedings of ACM CH1'95
Conference on Human Factors in Computing Systems, volume I . pp.
21W2217.1995.
[10] B. Smyth, P. Cotter and G. O'Hare, "Let's Get Personal Personalised
Television Listings on the Web, in 9th Irish Conference on Art$ cinl
Inrelligence and Conitive Science. Dublin, Ireland, 1998
[11] P. Resnick, N. lacovou, M. Suchak, P. Bergstorm and 1. Riedl, "Graup-
Lens: An Open Architecture for Collaborative Filtering of Nemews",
in Proceedings of ACM 1994 Conference on Computer Supported
Cooperative Work, pp. 175-186, Chapel Hill, North Carolina. 1994.
ACM.
[12] W. S. Lee, "Collaborative Learning for Recommender Systems", in Proc.
18th International Conf on Machine Learning. pp. 316321. Morgan
Kaufmann, San Francisco, CA, 2001
[13] G. Salton, A. Wong and C.S. Yang, A vector space model for automatic
indexing.
[14] Emmanouil Vozalis, Konstantinos G. Margaritis, Analysis of Recom-
mender Systems Algorithms
[15] Parsons, J., Ralph, P., & Gallagher K. (2004). Using viewing time to infer
user preference in recommender systems. AAAI Workshop in Se-
mantic Web Personalization, San Jose, California, July 2004
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 6, JUNE 2011, ISSN 2151-9617
HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/
WWW.JOURNALOFCOMPUTING.ORG 32

[16] Croft, W. Bruce. Advances in Information Retrieval: Recent Research
from the Center for Intelligent Information Retrieval. Hingham, MA,
USA: Kluwer Academic Publishers, 2000.
[17] G Salton, Automatic Text Processing: The transformation, analysis,
and retrieval of information by computer, Addison-Wesley, Massa-
chusetts, USA,1989
[18] Noriyuki Yamamoto, Mari Saito Mitsuhiro Miyazaki, Hiroyuki Koike,
Recommendation Algorithm Focused on Individual Viewpoints,
PAO Group, Intelligent Systems Research Laboratory, Information
Technologies Laboratories, Sony Corporation Tokyo, Japan.
[19] Cristina Gena and Liliana Ardissono, On the Construction of TV viewer
Stereotypes Starting from Lifestyles Surveys.
[20] Mark van Setten, Experiments with a recommendation technique that
learns category interests.

First A. Author Md.Ariful Islam Khandaker obtained Bsc Degree on
Computer Science & Engineering from Khulna University of Engi-
neering & Technology which is one of the top classes Engineering
Universitie of Bangladesh. Now he is persuing Msc Degree on Com-
puter Science & Engineering from another renowned university of
Bangladesh naemed Chittagong University of Engineerin & Technol-
ogy. He is now teaching in a university named International Islamic
University Chittagong, Bangladesh as a Lecturer in CSE Depart-
ment.He is working with Semantic Web now.

Second B. Author Abdul Kadar Muhammad Masum obtained Bsc
Degree on Computer Science & Engineering from International Is-
lamic University Chittagong, Bangladesh. He is now working as As-
sistant Professor in this university and doing Msc in CSE in United
International University of Bangladesh.

Second C. Author Golam Mostofa obtained Bsc Degree on Com-
puter Science & Engineering from International Islamic University
Chittagong, Bangladesh. He is now working as Assistant Professor in
Stampford University, Bangladesh in CSE Department.

You might also like