=Paper=
{{Paper
|id=Vol-1131/mindthegap14_3
|storemode=property
|title=When to Recommend What? A Study on the Role of Contextual Factors in IP-based TV Services
|pdfUrl=https://ceur-ws.org/Vol-1131/mindthegap14_3.pdf
|volume=Vol-1131
|dblpUrl=https://dblp.org/rec/conf/iconference/YuanSMH14
}}
==When to Recommend What? A Study on the Role of Contextual Factors in IP-based TV Services==
https://ceur-ws.org/Vol-1131/mindthegap14_3.pdf
When to Recommend What? A Study on the Role of
Contextual Factors in IP-based TV Services
Jing Yuan Fikret Sivrikaya Stefan Marx Frank Hopfgartner
Technische Universität Berlin
Ernst-Reuter-Platz 7, 10587 Berlin, Germany
{jing.yuan, fikret.sivrikaya, stefan.marx, frank.hopfgartner}@dai-labor.de
reached 87.2 million globally as of mid-2013. 17 mil-
lion of these subscribers have been acquired within the
Abstract last twelve months alone. This success story is due
to various factors. First of all, novel techniques for
Today’s IP-based TV services commonly the compression and streaming of multimedia content
strive for personalizing their content offers have been developed. Moreover, the rapid develop-
using complex recommendation systems to ment of fixed and mobile broadband communication
match their users’ interests. These systems technologies resulted in increased availability of band-
try to capture the relevance of content rec- width for the streaming of multimedia content over the
ommended to a user, which may also depend web. Apart from these technological advancements,
on many contextual factors such as time, lo- the main reason for the success of IPTV services can
cation, or social company. Nevertheless, in be considered as the flexible, dynamic access to content
most cases, these factors are either omitted or provided via these services. Differing from traditional
integrated in recommendation systems with- television channels that broadcast rather static con-
out a concrete modeling of what different roles tent for all consumers, content provided by IP-based
each may play on different users’ experiences. television and Video-on-Demand (VoD) services can
Do users really care about all of these spe- be adapted to the individual customer’s interests. An
cific factors? How do those factors interact important aspect of this adaptation process is the de-
with or influence each other? Can this inter- velopment of appropriate recommendation techniques
action be modeled commonly for all users or such as [2, 3, 4].
is it more specific to the user profile? To the These personalization techniques strongly depend
best of our knowledge, answers to these ques- on understanding users’ needs, which is, however, a
tions have not been studied in detail yet. In non-trivial task. Users’ needs and interests can change
this paper, we introduce the results of a ques- over time and can depend on external contextual fac-
tionnaire and a focus group discussion to elab- tors such as the time, location or company of other
orate on the influence of contextual factors on people [5]. Various studies (e.g., [6, 7, 8]) have shown
IP-based TV services from the users’ point-of- that recommender systems can benefit significantly
view. when these contextual factors are incorporated. Given
that we are all individuals though, it is not premature
1 Introduction to assume that contextual factors are not equally im-
portant for all of us. For example, the time of the day
According to a recent survey performed by Point Topic might be important for some people, but not so impor-
[1], the number of subscribers of IPTV services has tant for others. As far as we know, a detailed analysis
of the choice of contextual factors for a recommender
Copyright c 2014 for the individual papers by the paper’s au- has not been studied yet.
thors. Copying permitted for private and academic purposes.
This volume is published and copyrighted by its editors.
With this paper, we intend to shed some light on
In: U. Kruschwitz, F. Hopfgartner and C. Gurrin (eds.): Pro-
the role of contextual factors on individual users. The
ceedings of the MindTheGap’14 Workshop, Berlin, Germany, work is divided into two parts. First, we present the
4th March 2014, published at http://ceur-ws.org outcome of an online questionnaire where we asked
participants about their usage of IP-based TV ser- sumer of IP-based TV services. In this paper, we in-
vices. Second, we summarize the results of a focus clude social networks, which are a more detectable so-
group discussion where we discussed the results of the cial factor, into the discussion range to observe users’
questionnaire with various participants of our survey. attitudes towards social company.
The paper is structured as follows. In Section 2,
four possible influential contextual factors are pro- External breaking news may be treated not
posed, which form the basis of the questionnaire pre- only as a content type for recommendation, but also
sented and analyzed in Section 3. Section 4 summa- as a contextual factor. Studies in [15, 16] make use
rizes the participants’ opinions in the focus group dis- of trending topics on microblogs to mine real-time hot
cussion. Finally, Section 5 concludes the work and news. In IP-based TV services, external breaking news
outlines future work. might be quite relevant for the consumer’s choice on
programs beyond news, and thus worthy of being stud-
2 Contextual Factors ied as a contextual factor. For example, a famous
singer’s death may arouse users’ interests in his old
In literature, various contextual factors have been pro- music videos or concerts, apart from the news of his
posed that should be considered when providing rec- death itself.
ommendations. In the context of this paper, we con-
centrate on a subset of them, i.e., the most commonly In order to study the role of these contextual factors
used contextual factors. These include time, location, for individual users, we performed a user survey where
social company and external breaking news, each of we asked participants to answer specific questions on
which we elaborate next. five-point likert scale, multiple choice or radio. More-
over, we organized a focus group session where we dis-
Time Various researchers (e.g., [9, 10, 11]) focus
cussed the role of these factors with different types of
on time as contextual factor to improve their recom-
IP-based TV content consumers.
mendation algorithms. Although these works suggest
that time can be a very strong and thus helpful factor,
it remains unclear if this observation is valid for every 3 User Study Questionnaire
user in an IP-based TV scenario. For example, free-
In our design of the user questionnaire, we focused on
lancers with flexible working hours might not consider
the four specific contextual factors for IP-based TV
time to be a significant factor, while employees on a
recommendations introduced in the previous section.
strict work schedule generally consider it to be highly
At the beginning of the questionnaire, we gave users
important. We argue that further investigations are
a description of our “IP-based TV service” concept,
required to study the role of time as contextual factor.
which covers not just IPTV through set-top boxes,
Location is another frequently mentioned contex- but also WebTV and web-based mobile apps. In this
tual factor for recommenders [6, 12]. In most cases, lo- section we share some of the statistical results of the
cation as a contextual factor is considered by following survey, through which we try to provide a clearer pic-
simple matching rules. A concrete approach for build- ture of the contextual factors’ influence from the users’
ing relations between location and content evaluation perspective.
procedure has not been studied yet. Given the unclear
picture of location factor’s involvement in personaliza- 3.1 Respondents’ Basic Info
tion and recommendation methods, we consider it to
be a relevant factor that need to be investigated fur- The online questionnaire remained in effect through-
ther. out the month of August 2013, with a total of 51 re-
spondents. The demographic information of the ques-
Social company Users’ acceptance of social com- tionnaire respondents is listed in Table 1. All respon-
pany during IP-based TV service usage has been iden- dents are digital natives, i.e. were born after the start
tified in literature (e.g., [13, 14]). In a social con- of the digital age (around 1960); so their understand-
text, functionalities such as exchanging thoughts on ings of legacy TV (terrestrial, cable, satellite) and IP-
TV programs or recommending each other interest- based TV services are clear. Most participants are ei-
ing TV content are very common amongst users. So- ther employees or students at our university. In terms
cial context can provide users an opportunity to evade of their place of birth and residence, Asians and Euro-
the filtering bubble, which guides users to their own peans form the two largest groups of our respondents.
preferred directions, thus leading to large amount of This coincides with the survey result from Point Topic
hidden content. Nevertheless, it can not easily be as- [1], which shows that Asia and Europe are the two
sumed that social company is important for every con- biggest markets for IP-based TV content with 48.7%
and 36.6% market share of the worldwide IPTV sub- ests for specific types of content: “What kind of pro-
scribers, respectively. Given these similarities, we ar- grams would you prefer watching a) in the morning, b)
gue that our participants form a subset of the main during a break at daily work, c) in the evening and d)
target groups for such services. on weekends?” As presented in Figure 1, twelve basic
Table 1: Questionnaire Demographic Info genres of TV content were listed as choices for each
of the four categorical time periods. In line with the
intuitive reasoning, we observe the following trends:
i) weather report and daily news seem to be favor-
able choices in the morning or during a break at work,
when people usually spend much less time watching
TV; ii) similarly, during a break at work, those rela-
tively short TV content such as daily news, sport, mu-
sic and entertainment content are usually consumed;
iii) users’ preference in the evening and on weekends
show very similar behavior, with the comparatively
longer programs such as movies, TV series and doc-
umentaries outweighing other content types. Despite
of the resemblances to some TV company strategies,
users’ intuitive choices still make these trends worthy
Table 2 represents TV usage habits of the respon- of being referred to when recommending, especially for
dents. We first observe that a large majority con- VoD services.
sumes IP-based TV services much more than tradi-
tional TV, with more than half of the participants
spending at least five times more time on IP-based
TV services than on normal TV. Moreover, 86.3%
(19.6%+25.5%+41.2%) report that they have been us-
ing IP-based TV services for over two years. These
statistics confirm that the respondents to the ques-
tionnaire represent experienced IP-based TV service
users, possessing the required reference value for our
survey.
Table 2: Respondents Basic Usage Info
Figure 1: Users’ Choices on Program Categories
The next question that we cover is on the user’s
direct opinion on a more limited set of recommenda-
tion types given to them in a set of changing con-
texts regarding time and location: “Consider three
types of content recommendations provided to you at
the same time (habitual content at this time, breaking
news or events happening just now, friends’ instant
suggestions). In each of the contexts (at home in the
morning; at work hours during a break; at home in the
evening; on weekends), which of those recommended
contents are you most likely to choose for watching?”
As depicted in Figure 2, the users seem to be much
more interested in hearing about breaking news and
events during work hours or in the morning at home,
3.2 Context Influence on Content Selection
similar to the earlier question’s result. Conversely, the
The first question that we analyze in the questionnaire habitual content or friends’ suggestions become much
tries to capture the temporal changes in users’ inter- more favorable in the evening or on weekends. In other
words, the influence of contextual factors as break-
ing news and social company on users’ preference may
change with alterations in certain contexts as time and
location.
In addition to supporting the existence of contex-
tual factors’ influence on TV content selection or rec-
ommendations, which can be turned out from the first
question’s result, the second question’s result also pro-
Somewhat
vides an interesting insight on contextual factors’ mu- Not important
important
Can't decide Important Very important
tual influence, when we consider the breaking news
and social effects as contextual factors. Figure 3: Users’ Scoring Distribution on Contextual
Factors’ Importance
deviation for each of the factor’s ratings (Daily viewing
habit: 1.12, Time: 1.22, Location: 1.34, Social com-
pany: 1.20 and External breaking event: 1.32), we ob-
serve the largest variation also for Location, although
the difference among the four factors are again not so
significant. These spread distributions (according to
Chebyshev’s rule, there will be at least 3/4 of the data
within 2 standard deviations of the mean and at least
8/9 of the data within 3 standard deviations of the
mean) illustrate that there is no strong and unified ten-
Figure 2: Contextual Factors’ Influence on dency towards the valued importance of specific con-
Recommendation Type textual factors. Moreover, it seems that the contextual
factors’ influences are valued differently by the users,
and that there is no specific contextual factor that is
3.3 Users’ Perspective on the Importance of equally important for everyone. This further supports
Contextual Factors the existence of individual difference when considering
the importance of specific contextual factors, which
Aiming to study the importance of these contextual
differs from the usual overall consistent treatment of
factors from the individual users’ points of view, we
them.
posed the following question: “How important is each
of the following factors regarding their influence on We then compute the Pearson Correlation Coeffi-
your own (subjective) choice of TV programs recom- cient for each pair of the given factors based on re-
mended? ” Participants were then asked to assess the spondents’ ratings, as given in Figure 4. The fact that
importance of the four factors (Time, Location, So- the correlation between any two factors turns out to
cial company and External breaking events), in addi- be quite weak indicates that a user’s interpretation of
tion to the option of using their standard user profile each factor’s influence level may be independent on
(based on daily viewing habits). The assessment was their evaluation of the other factors.
based on a five-point likert scale, ranging from “1-Not
important, 2-Somewhat important, 3-Can’t decide, 4-
Important, 5-Very important”.
Figure 3 depicts the distribution of the respondents’ 0.43 0.14
ratings. Intuitively, scores for each factor’s importance 0.41
are relatively evenly distributed in the five-point lik- 0.41
0.24
ert scale. Considering the average score for each fac-
tor’s importance (Daily Viewing Habit: 2.78, Time: 0.43
0.40 0.16
2.98, Location: 3.39, Social company: 3.25 and Exter- 0.14
nal breaking events: 3.29), location is viewed among
Breaking Events
respondents as the one factor having slightly more in- 0.48
fluence than others on TV content selection, although
there is no clear winner. Obviously, there is no evi- Figure 4: Pearson Correlation Coefficient of Different
dent preference for any specific contextual factor from Pair of Factors
a general view. When analyzing the sample standard
3.4 Subjective Responses paring them as follows: “Location means more than
timing in my case. I only watch program through in-
We also provided two free-text style opinion questions
ternet at home. Whenever in the office, I’ll be busy
in the questionnaire to gain additional user insight on
with my work and won’t open any TV related applica-
the assessment of IP-based TV services and contextual
tions.” Along the same lines, P1 expressed his view as:
factors.
“Whether I’m working or on vacation will result dif-
Q1: From your point of view, what features should ferently on my willingness to accept recommendations.
a perfect IP-based TV service offer? For me, time of the day, day of the week can regularly
determine my status of busy or not and thereby drive
Q2: What other contextual factors may influence
my choices.” P6 shared a different perspective from
your decision to follow a certain program on IP-
her own experience: “Whenever it is or wherever I
based TV services?
am, if I am using IP-based TV services for recreation,
Even though these questions did not have any pre- it means I have time and will enjoy the content I’m
scribed options for the answers, the responses have interested in; so both time and location factors won’t
shown some natural clustering around a few concepts. influence that much.”
For Q1, several respondents explicitly referred to Concluding from these statements, we argue that
the contextual factors External breaking news and “busy or not” might be a decisive factor for users’
events and Social company as requirements for an ideal choices when watching TV content, while the directly
IP-based TV service. In addition, more variety in con- measurable contextual factors time and location are
tent, less advertisement, free of charge service, and a less important but might be clues to figure out users’
clear and fast UI were also suggested. Responses to status of “busy or not”.
Q2 supplemented influential contextual factors with
some inspiring comments. Some claimed that the sta-
tus of “busy or not” would play a central role in users’ 4.1.2 Comments on Social Company
preferred type and length of TV content, and others
suggested mood as an independent contextual factor. When the topic moved on to social company, P2 and
As implicit factors though, “busy or not” and “mood” P8 have shown strong interest by expressing that it is
can not be so easily detected from existing datasets. always a great experience to exchange thoughts with
Therefore, any clue that can help deduce users’ such friends on the programs of common interest, and that
status would be quite valuable. Some respondents also they always get great suggestions from friends. P2
indicated that the quality of TV programs is of impor- mentioned the more concrete case of watching soc-
tance. Comprehensive factors such as players, guests, cer games while discussing and sharing opinions with
theme were all referred to as quality evaluation indica- friends. Yet, on the totally opposite side, P3 com-
tors on programs, which are already well studied and plained, “I really hate being bothered by others; I just
integrated in most recommendation systems. would like to be immersed in my own interested pro-
grams alone.” These statements illustrate that content
4 Focus Group Discussion categories and occasions should be carefully considered
when providing social-based recommendation.
In order to further find out users’ personal usage expe-
riences and remedy their unavoidable misunderstand-
ings of objective questions, we invited participants of 4.1.3 Being Cautious with Breaking News and
our questionnaire to join a focus group discussion. Events
Eight respondents were able to participate in the dis-
cussion session. In this section, we label these par- As to the newly proposed contextual factor of break-
ticipants as P1 to P8 to share some of their valuable ing news and events, P8 stated, “I feel disturbed when
ideas. small windows pop-up to remind me of some so-called
’news’, unless they are really appealing to me.” P7
4.1 Discussion on Contextual Factors supported this with: “I routinely view news from the
news websites; I don’t think it’s necessary to get rec-
Considering the main purpose of the work in this pa- ommendation from a TV application with respect to
per, participants’ opinions on contextual factors were news again.” Just as the participants warned, TV rec-
undoubtedly the main focus group discussion. ommenders should be extra cautious in the way they
select and notify their users of breaking news and
4.1.1 Time and Location vs. “Busy or Not”
events. Otherwise the recommendation could be more
Contextual factors of time and location were always annoying than appealing, no matter how important it
referred together. Participant 3 (P3 ) started by com- is.
4.2 Other Points of Discussion In accordance with the guides and insights turned out
by this paper, our next step is to design and develop
Aside from the intended discussion on proposed con-
a context-adaptive recommender system for our own
textual factors, there were other points initiated by
IP-based TV service that incorporates these factors.
participants.
4.2.1 Users’ Sensitivity to Content Quality Acknowledgements
Opinions on popularity and quality of TV content, as The first author has been funded by the Chinese Schol-
appeared in subjective responses of the questionnaire, arship Council.
were proposed again in the focus group. P4 stated,
“Popularity is a useful reference when I choose TV References
content, while it won’t work sometimes since content’s
[1] Point topic iptv statistics - market analysis q2
popularity can’t directly determine its quality.” P2
2013, June 2013.
continued, “I also found that some so-called popular
TV content are pushed in front of us only due to com- [2] James Davidson, Benjamin Liebald, Junning Liu,
mercial reasons rather than users’ preference.” Appar- Palash Nandy, Taylor Van Vleet, Ullas Gargi,
ently, users are more sensitive to TV content’s quality Sujoy Gupta, Yu He, Mike Lambert, Blake Liv-
now than ever before, and they won’t be satisfied with ingston, and Dasarathi Sampath. The youtube
just the popularity statistics. Thus distinguishing high video recommendation system. In Proceedings of
quality content would be quite an important aspect for the fourth ACM conference on Recommender sys-
recommenders. tems, RecSys ’10, pages 293–296, New York, NY,
USA, 2010. ACM.
4.2.2 Bookmarking also Implies Success for
Recommendations [3] Yu Xin and Harald Steck. Multi-value proba-
Another unexpected acquirement from the focus group bilistic matrix factorization for ip-tv recommen-
was users’ supplementary view on recommender’s ef- dations. In Proceedings of the fifth ACM confer-
fectiveness. P1 said, “When I don’t have time to ence on Recommender systems, RecSys ’11, pages
watch TV content that was correctly recommended, 221–228, New York, NY, USA, 2011. ACM.
I’ll bookmark it and watch it later on.” P2 commented [4] Yu-Cheng Chen, Hsien-Chao Huang, and Yueh-
similarly: “I also have the same habit of bookmarking Min Huang. Community-based program recom-
pages when I use WebTV; it is quite convenient.” Ev- mendation for the next generation electronic pro-
idently, apart from users’ instant positive reaction to gram guide. Consumer Electronics, IEEE Trans-
recommendations, such as clicking or watching dura- actions on, 55(2):707 –712, may 2009.
tion, the behavior of bookmarking can also be a rep-
resentative indicator for a recommender’s success. [5] Gediminas Adomavicius and Alexander Tuzhilin.
Context-aware recommender systems. In Pro-
5 Conclusion and Future Work ceedings of the 2008 ACM conference on Recom-
mender systems, RecSys’08, pages 335–336, New
In this paper, we analyzed the role of common con-
York, NY, USA, 2008. ACM.
textual factors that are usually applied to recommend
content to users of IP-based TV services. We ad- [6] Songbo Song, H. Moustafa, and H. Afifi. Ad-
dressed this question from the perspective of the con- vanced iptv services personalization through
sumers, i.e., we asked for their opinions in an online context-aware content recommendation. Multi-
questionnaire and a succeeding focus group discus- media, IEEE Transactions on, 14(6):1528 –1537,
sion. We conclude from our samples that i) contex- dec. 2012.
tual factors’ influence and their interplay indeed ex-
ist; ii) users’ attitudes toward contextual factors’ in- [7] Paolo Cremonesi and Roberto Turrin. Time-
fluence are individually different, which refutes tradi- evolution of iptv recommender systems. In Pro-
tional contextual factors’ treatments of modeling them ceedings of the 8th international interactive con-
separately or incorporating them equally on every in- ference on Interactive TV & Video, EuroITV ’10,
dividual. At the same time, users’ new suggestions on pages 105–114, New York, NY, USA, 2010. ACM.
contextual factors such as “busy or not” and mood,
their advanced cognition of TV content’s quality and [8] Alexandros Karatzoglou, Xavier Amatriain, Linas
their taboos of being interrupted by dull recommen- Baltrunas, and Nuria Oliver. Multiverse recom-
dations are all factors that should be studied further. mendation: n-dimensional tensor factorization for
context-aware collaborative filtering. In Proceed- [13] Keith Mitchell, Andrew Jones, Johnathan Ish-
ings of the fourth ACM conference on Recom- mael, and Nicholas J. P. Race. Social tv: The
mender systems, RecSys ’10, pages 79–86, New impact of social awareness on content naviga-
York, NY, USA, 2010. ACM. tion within iptv systems. Comput. Entertain.,
9(3):19:1–19:29, November 2011.
[9] Yehuda Koren, Robert Bell, and Chris Volinsky.
Matrix factorization techniques for recommender [14] Daniel Schreiber. Social iptv: a survey on chances
systems. Computer, 42(8):30–37, August 2009. and user-acceptance. In Martin Atzmüler, Do-
minik Benz, Andreas Hotho, and Gerd Stumme,
[10] Yi Ding and Xue Li. Time weight collaborative editors, Proceedings of LWA2010 - Workshop-
filtering. In Proceedings of the 14th ACM interna- Woche: Lernen, Wissen & Adaptivitaet, Kassel,
tional conference on Information and knowledge Germany, 2010.
management, CIKM ’05, pages 485–492, New
[15] Owen Phelan, Kevin McCarthy, and Barry
York, NY, USA, 2005. ACM.
Smyth. Using twitter to recommend real-time
topical news. In Proceedings of the third ACM
[11] Pedro G. Campos, Fernando Dı́ez, and Iván Can-
conference on Recommender systems, RecSys ’09,
tador. Time-aware recommender systems: a com-
pages 385–388, New York, NY, USA, 2009. ACM.
prehensive survey and analysis of existing evalua-
tion protocols. User Model. User-Adapt. Interact., [16] Nargis Pervin, Fang Fang, Anindya Datta,
24(1-2):67–119, 2014. Kaushik Dutta, and Debra Vandermeer. Fast,
scalable, and context-sensitive detection of trend-
[12] M. Dabrowski, J. Gromada, and H. Moustafa. ing topics in microblog post streams. ACM
Context-awareness for iptv services personaliza- Trans. Manage. Inf. Syst., 3(4):19:1–19:24, Jan-
tion. In Innovative Mobile and Internet Services uary 2013.
in Ubiquitous Computing (IMIS), 2012 Sixth In-
ternational Conference on, pages 37–44, July.