Icete content-based filtering with applications on tv viewing data

CONTENT-BASED FILTERING WITH APPLICATION ON TV
VIEWING DATA
Preparation of Camera-Ready Contributions to INSTICC Proceedings

Elaine Cecília Gatto, Sergio Donizetti Zorzo
Department of Computer Science, Federal University of São Carlos,
Rodovia Washington Luís, Km 235, PO Box 676, São Carlos, Brazil
elaine_gatto@dc.ufscar.br, zorzo@dc.ufscar.br

Keywords: Personalization, Recommendation, Information Filtering, Brazilian Digital TV, Content-Based Filtering,
Recommendation System, Collaborative Filtering, Hybrid Filtering, One-Seg, Full-Seg, Middleware Ginga,
Cosine, Apriori.

Abstract: Recommendation systems provide recommendation based on information about users’ preferences.
Information Filtering is used by recommendation systems so as information can be processed and suggested
to users; and Content-Based Filtering is an Information Filtering approach very used in recommendation
systems. Content-Based Filtering analyses the correlation of items content with the user’s profile,
suggesting relevant items and putting away irrelevant items. Recommendation systems, which are very
much used on the Internet, have been studied in order to be used on Digital TV context, and there already
are several works in this sense. As they are used on the Internet, recommendation systems can be used in
Digital TV in order to recommend TV programs, publicity and advertisement and also the electronic
commerce. Thus, within Digital TV context, the items can be programs, advertisements and the products to
be sold; and using Content-Based Filtering in the recommendation programs, for instance, these programs’
contents can be correlated with the user’s preferences, which in this scenario, are the type of program one
wants to watch. This paper presents the studies accomplished with Content-Based Filtering with application
on Digital TV data. The survey aims at observing and evaluating how some filtering techniques based on
content can be used in recommendation systems in Digital TV context.

1 INTRODUCTION (Bozios et al, 2001), (Gutta et al, 2000), (Das and
Horst, 1998), among others.
Digital TV implementation in Brazil provides Recommendation systems can contribute to a
new markets which can be explored. Well-succeeded better use of Digital TV in residences, in groups or
technologies as those in Web environment, for individually, in a cell phone, for example. These
example, can be applied in Digital TV domain and systems can help the user to choose the program,
achieve the same success. avoiding waste of time and of course, suggesting to
The interaction either through the remote control the user programs which really interest him.
or the cell phone keyboard etc by the user today, will Moreover, recommendation systems can be applied
allow many applications to be carried to this to publicity and advertisement on Digital TV, as
environment. well as in the T-Commerce.
One of the areas which has been extensively This paper is structured as follows: Section 1
studied and is well-succeeded in the Web is that of provides a brief introduction to the survey, Section 2
personalization. There are some surveys concerning deals briefly with recommendation systems and its
recommendation systems for Digital TV as for techniques; Section 3 quickly describes Brazilian
example (Ávila, 2010), (Lucas, 2009), (Uribe, current conditions related to Digital TV; Section 4
2009), (Solla et al, 2008), (Bar et al, 2008), presents tests performed with TV viewing data;
(Einarsson, 2007), (Chorianopoulus, 2007), (Choi, Section 5 presents the outcomes from the tests and
Koh and Lee, 2007), (Yu et al, 2006), (Silva, 2005), Section 6 concludes the paper.

2 RECOMMENDER SYSTEMS Association rules interconnect objects trying to
present characteristics and tendencies. Association
In a typical recommendation system, the users findings must evidence either common associations
provide recommendation as inputs which are then or uncommon associations.
added and directed to proper receivers. (Resnick, Apriori algoruthm is frequently used to mine
1997) association rules. Apriori operates with a high
With the first articles on collaborative filtering number of attributes, creating several combinations
around the 90’s, recommendation systems became among them and performing consecutive search in
an important area of research. Recommendation the whole database, keeping a great performance in
systems comprise several technologies as cognitive terms of time spent in the processing.
science, approximation theory, information The algorithim tries to find all the relevant
recovery, forecast theories, among others, and can association rules between the items, which have the
be applied to several domains. X format (precedent) ==> Y (consequent). If x% of
The recommendation problem in its most transactions which have X also have Y, so x%
common form is reduced to a way of evaluating represents the confidence factor (power of
items which were not seen by a user. Evaluation of confidence of the rule). The support factor is a
non-evaluated items can be estimated in many measure corresponding to x% of X and Y occurance
different ways, frequently classified according to its simultaneously upon the total of registers
approach to classification estimate. In Sections 2.1, (frequency). (Witten, 2005)
2.2 and 2.3, recommendation systems classification
is presented. (Adomavicius, 2005) 2.2.2 Cossine
Cosine is a similarity measure, a metrics which can
2.1 Content-Based Filtering be applied to discover if an item has correlation or
not with the user profile. In many recommendation
Content-Based Filtering (CBF) uses the content systems for the Web, the applied techniques use the
attributes to describe the content of the items and evaluation performed by the users, for the products
then calculate the similarity. This approach does not consumed to calculate the similarity.
depend on other users’ evaluation about the items. In our context, this evaluation by the user is not
(Einarsson, 2007) possible yet, therefore, we used the time a person
CBF is an information recovery technique spent watching the program as an evaluation. In the
which bases its forecast on the fact that previous same way we found an alternative, virtual stores
preferences of the users are reliable indicators for which do not require users’ evaluation for its
future behavior. (Chorianopoulos, 2007) products can consider “consumed product” and not
In order to formulate recommendations, a “non-consumed product” as an evaluation.
variety of algorithms has been proposed to evaluate A binary vector is a set of two elements, x and y.
the content of documents and find regularities. Some In an n-dimensional space, where n is the number of
of these algorithms operate with classification items of the vector, it is possible; therefore, calculate
knowledge and others operate with the problem of the cosine between the vectors, thus evaluating the
regression. (Pazani, 1999) similarity between the user profile and its history.
Some of the problems and limitations found The similarity is high when the cosine value is high.
in systems using CBF are super specialization, the The cosine formula is presented below:
problem of the new user and the analyses of limited
content. The following 2.2.1 and 2.2.2 subsections 
  ( p.e )
describe two techniques which can be used in CBF cos( p, e )    (1)
and which were applied in our survey. | p |.| e |
(Adomavicius, 2005)
Where is the profile vector and is the EPG
2.2.1 Apriori vector. The symbol means the profile vector
standard and the symbol the EPG vector standard .
The algorithms of association techniques identify
(Torres, 2004, 2009)
associations between register of data related in some
way. The major premise finds elements which
require the presence of others in a same transaction,
aiming at determining what is related.

2.2 Collaborative Filtering Table 1: Number of Individuals per Residence.
Residence 1 2 3 4 5 6
Collaborative Filtering (CF) is a technique which Individuals 2 3 3 2 2 3
uses the similarity between users in order to generate
TVs 1 1 2 2 1 2
recommendations and it first came to light in the
90’s, with Tapestry system, different from CBF
Table 2: Social-economic characteristics at Residences 1,
which calculates the similarity between the items. 2 and 3.
CF stores the users’ evaluation about each item
and based on this information, finds people with Residence 1 2 3
similar profile, the so-called nearest neighbors, who Social
DE C C
are then gathered and the products with high Class
evaluations by neighbors are recommended. Residence 1 2 3
Age of the
(Balabanovic, 1997; Torres, 2004) 44 45 39
hostess
Even solving some CBF problems, CF introduce Level of
others as the problem of the new user, the problem Incomplete Incomplete Incomplete
education of
of the new item and the sparcity. Primary High High
the owner of
School School School
the house
2.3 Hybrid Filtering Individual 1
Female Female Female
gender
Hybrid filtering mixes CBF and CF in a sole system, Individual 1
8 48 40
improving recommendation offered to user and thus, age
seeks to solve some of the problems introduced by Individual 2
Female Male Male
gender
both techniques.
Individual 2
This way, recommendation methods in this - 17 13
age
category can be matched in many ways: a) CF Individual 3
sequentially processed after CBF; CBF sequentially - Female Female
gender
processed after CF and CBF parallelly processed Individual 3
with the CF. (Einarsson, 2007; Adomavicius, 2005) - - -
age

Table 3: Social-economic characteristics in Residences 4,
3 BRAZILIAN DTV 5 and 6.
Residence 4 5 6
Since December, 2007 in Brazil, the implantation Social
AB C AB
of Brazilian Digital TV has been innovating by Class
matching Japanese technology with technology Age of the
32 60 36
developed by Brazilian universities. hostess
Besides having all the advantages of Japanese Level of
Complete Complete Complete
education of
system, Brazilian system counts on Ginga High High High
the owner of
Middleware which uses LUA, NCL and Java School School School
the house
languages, totally developed by national researches. Individual 1
Peru, Argentina, Chile and Venezuela chose the Female Female Female
gender
Nipo-Brazilian standard of Digital TV which is Individual 1
already part of UIT. Nipo-Brazilian standard offers 30 77 38
age
quality of image and sound, mobility, portability, Individual 2
Male Male Male
flexible interactivity; it is free of royalties and gender
provides the development of commercial, playful, Individual 2
- - 14
informative, governmental, social inclusion age
applications, among others. (SBTVD Forum, 2009) Individual 3
- - Male
gender
The standard (ABNT NBR 1564, 2008) defines
Individual 3
the set of essential functionalities required from - - -
age
reception devices of 13-segment digital television –
Full-seg – as well as from one-segment – One-seg –
designated to receive signals in fix, mobile and
portable modality.

Still according to this standard, full-seg presents the names of broadcasting stations with the
classification is applicable to digital converters – set- number of programs and genres transmitted.
top box – and to 13-segment receptors integrated to
the viewing screen, but not exclusive to these; and
one-seg classification is designated to portable-type
receptors – handheld – specially recommended for
smaller screens, commonly up to 17,80 inches.
The content can be then displayed in many
different devices, as well as diversified services can
also be formulated for each one, allowing the
creating of new business models and new
opportunities for professionals.
Ginga is the name of the middleware developed
by researches performed by Telemedia laboratories
at PUC-Rio and LAViD at UFPB. The middleware
is divided in Ginga-NCL/LUA, corresponding to the
declarative part and Ginga-J, the imperative part.
(GINGA, 2010)

4 TESTS Figure 1: Types of data composing EPG.

So as the test could be performed, data Table 4: Number of broadcasting stations, programs and
genres in EPG.
corresponding to TV viewing and from the TV guide
were used. This data was provided by IBOPE. The Broadcasting
Programs Genres/Subgenres
characteristics of this data and the performed tests stations
are detailed in the following subsections. 1 Bandeirantes 70 23
2 Gazeta 40 10
4.1 Characteristics of Residence 3 Globo 76 18
4 MTV 149 12
5 RBI TV 46 12
Data provided by IBOPE correspond to 15-day
6 Record 42 15
monitoring at 6 Brazilian residences with Open TV
7 Record News 100 10
programs. 8 Rede TV 67 20
These residences were monitored minute-to- 9 SBT 61 15
minute, as well as each individual was monitored 10 TV Cultura 167 22
separately. Table 1 shows the number of individuals
and TVs by residence, Table 2 presents the social- 4.2.2 User History
economic information of residences 1, 2 and 3; and
Table 3 deals with residences 4, 5 and 6. Users’ viewing history is necessary in order to
discover their preferences.
4.2 Characteristics of Date In the Digital TV context we are considering,
this data are collected and stored implicitly.
Data used for these tests undergone a manual Figure 2 presents the composition of data and
process of adaptation. For each of the algorithms Table 3 presents a sample of data in the viewing
used, it was necessary a pre manual processing so as history.
they could be correctly analyzed and used.
Subsections 4.2.1 and 4.2.2 detail the composition of Table 5: Amostra do histórico de usuário.
these data.
Field Content
startSyntonization 2008-03-05 09:28:00
4.2.1 EPG endSyntonization 2008-03-05 12:59:00
durationSyntonization 03:31:00
EPG provided by IBOPE corresponds to the 15-day
Date 2008-03-05
schedule of 10 broadcasting stations. Figure 1 shows
timeStart 09:28:00
the types of data which composes EPG and Table 4 timeEnd 12:59:00

duration 211 marked in the matrix with the value of 1 and the
periodSyntonization morning remaining is marked with the value of 0. This has
day of the week Wednesday been done for all programs composing EPG.
Programcode 003217 After this, a table called “profile” was created
Programname HOJE EM DIA which stores the user profile found consulting SQL,
Broadcastingstationcode 006 which is showed in a simple way, in Figure 3 below.
Broadcastingstationname Record The “profile” table is presented in Figure 4.
Genre 0x6
Genredescriber Variety Select avg(ded1), avg(dee1), …,
Subgenre 0X0F avg(vs1)
Subgenredescriber Others from (select domicilio.nomePrograma,
genreSubgenre 0x6_0X0F domicilio.descritorGeneroSubgenero,
GeneroSubgenerodescriber Variety_Others duracao*DED as ded1,
duracao*DEE as dee1, …,
duracao*VS as vs1
from domicilio, matrizepg
where domicilio.nomePrograma =
matrizepg.nomePrograma
order by duracao desc) as result;

After that, a variable was set:

set @profilenorm=
(select sqrt(ded1*ded1+dee1*dee1+ …
+vs1*vs1)from profile);

Figure 2: Types of data composing the user history.

4.3 Methodology
In order to carry out the tests, we simulated the
generation of recommendations and profile for each
residence, using two different techniques, Apriori
and Cosine. Figure 3: Fields added to EPG generating EPG
For the Cosine, we used MySql databank. For Matrix.
each new day, we inserted in the databank
correspondent to the viewings and then, we applied
the recommendation technique, we discovered the
profile and which program to recommend.
The process occurs in an interactive systematic
way. First, data corresponding to the first day of
monitoring is inserted in the databank and the EPG
matrix is created, that is, EPG is transformed in a
matrix containing, besides the data in Figure 2, the
Genres and Subgenres of each program separately,
as presented in Figure 3. Each abbreviation indicates
one genre/subgenre.
If a program belongs to one or more
genre/subgenre, as for example, sport and Figure 4: Table Profile.
documentary journalism, these genres/subgenres are

And finally, the final result with the following For the case of Cosine, the existence of programs
consult: seen by the user in the following day in the results
based in the previous day was verified. This was the
select nomePrograma, best way for the evaluation, for the evaluation
descritorGeneroSubgenero, cannot be done directly with the users, however, it is
dot/(@profilenorm*norm) as cos, possible to know what the user has seen before and
DED, DEE, …, VS
from (select nomePrograma,
after each step.
descritorGeneroSubgenero, Thus, two additional tables were created; one in
sqrt(DED*DED+DEE*DEE+…+VS*VS) as norm, order to store the result of the cosine and the other to
DED*ded1+DEE*dee1+…+VS*vs1) as dot, store only what was seen in the following day. These
DED, DEE, …, VS tables were called “recommend” and
from matrizepg, profile) as normdot “residence_test” and the following SQL consult was
group by nomePrograma used to evaluate:
order by cos asc;
select r.*, dt.nomePrograma,
Thus, the programs which can be recommended dt.descritorGeneroSubgenero
to the user according to his profile were found. The from recomenda r, domicilio_teste dt
same thing can be done to fid only the where dt.nomePrograma = r.nomePrograma
genres/subgenres. group by r.nomePrograma
For Apriori, Weka tool was used having as order by cos desc;
parameters minima support o,1, reliance 0,9, class
attribute index -1, total of 20 rules and enabled car This way it is possible to discover if in the
providing the mining of the association rules instead following day, the individual watched some program
general rules of association. which is in the “recommend” table and to verify the
StringToNominal and NumericToNominal value of its cosine. If this value is near 1, then we
conversion filters were also applied in some fields, can say that the cosine gave a right forecast.
generating the rules and saving the outputs. Below is A behavior in which 5 recommendations were
a small sample of these rules: offered was simulated. If any of these 5
recommendations were seen on the next day and if
1.genero=0x62==>descritor=Variedade_Out its cosine is near 1, so it is assumed that the
ros2conf:(1) recommendation was accepted.
Figures 5 to 10 present the percentage of right
2.descGenero=Variedade2==>descritor=Var cosine, during 15 days of monitoring in each
iedade_Outros2conf:(1) residence, according to our methodology of
3.subGenero=0X0F2==>descritor=Variedade
simulation. Figure 11 presents the average of all
_Outros2conf:(1) residences.
Graphics were generated with the following
4.descSubGenero=Outros2==>descritor=Var formula:
iedade_Outros2conf:(1)

5.genSubg=0x6_0X0F2==>descritor=Varieda Number of Hits (0 a 5)
de_Outros2conf:(1) Percentage= (2)
Number of recommendations (5)
6.dia=2008-03-
05genero=0x62==>descritor=Variedade_Out
ros2conf:(1) For the case of Apriori, it was possible to verify
if the individual had seen some of the
genres/subgenres identified in the rules in the
following day. These are a little different approach.
5 RESULTS While in Cosine the operation was direct with the
names of the programs, in Apriori, the genres and its
After describing the methodology used, this sections respective subgenres were used.
presents the results. The techniques were applied; The same methodology to simulate the cosine
the results were evaluated and verified to see if was used for the Apriori. Figures from 12 to 17
correct recommendation was being generated. present the hit percentage of Apriori, during 15 days
of monitoring in each residence, according to the

simulation methodology. Figure 18 presents the
average of all residences and Figure 19 presents a
comparison between the averages of each one of the
techniques for all the residences.

Figura 8: Percentage of cosine hits, during 15 days in
residence 4.

residence 1.

residence 5.

residence 2.

residence 6.

residence 3.

Figura 11: Average of the Cosine in all residences. Figura 14: Percentage of Apriori hits, during 15 days,
in residence 3.

Figura 12: Percentage of Apriori hits, during 15 days, Figura 15: Percentage of Apriori hits, during 15 days,
in residence 1. in residence 4.

Figura 13: Percentage of Apriori hits, during 15 days, Figura 16: Percentage of Apriori hits, during 15 days,
in residence 2. in residence 5.

Table 6: Difference between Apriori and Cosine.
Residence 1 19%
Residence 2 8%
Residence 3 5%
Residence 4 16%
Residence 5 8%
Residence 6 28%

However, apriori provided other kinds of
information which are difficult to collect with the
cosine, concerning the user's behavior in each
Figura 17: Percentage of Apriori hits, during 15 days,
residence. While cosine is focused to select the
in residence 6. programs to be recommended according to the
profile generated also by the cosine, it is possible to
use apriori to find out other characteristics and thus
improve the quality of recommendations.
Table 7 present some of these characteristics.
This table presents the day of the week, the period of
the day, the genre/subgenre and the broadcasting
station watched by each one in the residences. This
information is independent, for example, a residence
might have watched soap opera, but this soap opera
is not necessarily from the most watched
broadcasting station

Table 7: Characteristics found out with apriori.
Figura 18: Average of the Apriori in all residences.
Period Broadca
Day of the Genre/
R of the sting
week Subgenre
day station
Afterno Variety_other
1 Thursday record
on s
Wednesda Soap Opera_
2 Evening Globo
y Soap Opera
children_child
3 Thursday Evening Globo
ren
Soap Opera_
4 Sunday Evening Record
Soap Opera
Journalism_ne
5 Friday Evening Globo
wcast
Soap Opera_
6 Friday Evening Record
Soap Opera
Figura 19: Comparison of the hits average between
Apriori and the Cosine in all residences.
It could also be seen that the apriori used in these
data tend to be super-specialized, always finding the
Certainly, the difference between the techniques same genres and subgenres to recommend. This
is visible and presented in Table 6. It is important to shows that it is necessary operate together with other
point out that although the methodology is the same techniques to create the surprise recommendation to
for both, the techniques were observed and analyzed the user, particularly in this case.
by different point of views, the cosine directed to the The data we have are simple and do not have
name of the program and the apriori for details as synopsis, name of the actors, directors,
genres/subgenres. sport categories etc. It is expected that, in Brazilian
Digital TV, these attributes are present, increasing
the probabilities of recommending not only the
obvious but also something new that the user would
probably watch.

6 CONCLUSION Proceedings of the Technological Forecasting &
Social Change, p. 1043-1053, 2007.
Chorianopoulos, K. Personalized and mobile digital TV
According to the studies presented herein, it is applications. In Proceedings of the Multimedia Tools
possible to apply FBC in TV viewing data and thus, and Aplications, p. 1- 10, vol.36, 27 January 2007.
it can also be applied for developing Cristo, M. Sistemas de Recomendação, Métodos e
recommendation systems for Digital TV, Avaliação. 81 slides. 2009.
particularly in Brazil. Das, D. and ter Horst, H. Recommder Systems for TV. In
Two different techniques were used in the same Proceedings of 15 th AAAI Conference, Madison,
data and it was possible to note that, despite the Wisconsin, July 1998.
Einarsson, O. P. Content Personalization for Mobile TV
differences among them, both can be used in order
Combining Content-Based and Collavorative Filtering.
to find out the profile and to provide Master Thesis. Center for Information and
recommendations, as well as they can be used Communication Technologies. Technical Univesity of
together to provide even better recommendations. Denmark. August 22, 2007.
There are also other FBC and FC techniques Fórum SBTVD. TV digital nipo-brasileira agora é
which will be tested in future works, together with oficialmente referência mundial. Available in:
hybrid techniques. More detailed data is also <http://www.forumsbtvd.org.br/materias.asp?id=238>,
expected as synopsis, indicative classification, Acess in January 2010, 09h00.
among others, in order to improve the quality of Ginga. Available in: <http://www.ginga.org.br/>, Acess in
January 2010. http://www.ginga.org.br/
recommendations in TV viewing domain.
Gutta, S. et al. TV Content Recommender System. In
Proceedings of the 17th National Conference of AAAI,
Austin, TX, 2000.
ACKNOWLEDGEMENTS Lucas, A. Personalização para Televisão Digital utilizando
a estratégia de Sistema de Recomendação para
ambientes multiusuário. 103 pages. 2009.
We thank IBOPE for providing real data about the
Pazzani, M. J. A framework for Collaborative, Content-
electronic program guide and also the viewer’s Based and Demographic Filtering. Artificial
behavior data from March, 05, 2008 to March, 19, Intelligence Review, p. 393-408, December 1999.
2008. Resnick, P.; Varian, H. R. Recommender Systems.
Communications of the ACM, New York, vol. 40, n. 3,
p. 77-87, March 1997.
Silva, F. S. Personalização de Conteúdo na TVDI Através
REFERENCES de um Sistema de Recomendação Personalizada de
Programas de TV (SRPTV). In III Fórum de
ABNT NBR 15604:2008. Televisão Digital Terreste Oportunidades em Televisão Digital Interativa, Poços
Receptores. 07/04/2008. 68 pages. de Caldas, Minas Gerais, Brasil, 2005.
Ávila, P. M. Recommender TV: Suporte ao Solla, A. G. et al. ZapTV: Personalized User-Generated
Desenvolvimento de Aplicações de Recomendação Content for Handheld Devices in DVB-H Mobile
para o Sistema Brasileiro de TV Digital. Dissertação Newtorks. In Proceedings 6th European Interactive
de Mestrado. 90 pages, 2010. TV Conference, p.193-203, Salzburg, Áustria, 03-
Adomavicius, G.; Tuzhilin, A. Towards the Next 04/07/2008.
Generation of Recommenders Systems: A Survey of Torres, R. Personalização na Internet. Editora Novatec,
the State-of-the-Art and Possible Extensions. IEEE 157 pages, São Paulo, Brazil, 2004.
Transactions on Knowledge and Data Engineering, Uribe, S. et al. Mobile TV Targeted Advertisement and
vol. 17, Issue 6, p. 734-749, June 2005. Content Personalization. In 16th International
Balabanovic, M. ; Shohan, Y. Content-Based, Workshop Conference on Systems, Signals and Image
Collaborative Recommendation. Communications of Processing, Chalkida, Greece, 18-19/06/2009.
the ACM, New York, vol. 40, n.3, p. 66-72, March Witten, I. H.; Frank, E. Data Mining: Practical Machine
1997. Learning Tools and Techniques, 2nd Edition, Morgan
Bär, A. et al. A Lightweight Mobile TV Recommender: Kaufmann, 525 pages, June 2005.
Towards a One-Click-to-Watch Experience. In Yu, Z. et al. TV program recommendation for multiple
Proceedings 6th European Interactive TV Conference, viewers based on user profile merging. In Proceedings
p.142-147, Salzburg, Áustria, 03-04/07/2008. of the User Model User-Adap Inter, p. 63-82, 2006.
Bozios, T. et al. Advanced Techniques for Personalized
Advertising in a Digital TV Environment: The iMedia
System. In Proceedings of the eBusiness and eWork
Conference, p. 1025-1031, IOS press, 2001.
Choi, J. Y.; Koh, D.; Lee, J. Ex-ante simulation of mobile
TV market based on consumers’ preference data. In

Icete content-based filtering with applications on tv viewing data

More Related Content

What's hot (17)

Viewers also liked (6)

Similar to Icete content-based filtering with applications on tv viewing data (20)

More from Elaine Cecília Gatto (20)

Recently uploaded (20)

Icete content-based filtering with applications on tv viewing data