Convergence of Web and TV Broadcast Data for
Adaptive Content Access and Navigation
Pieter Bellekens, Kees van der Sluijs, Lora Aroyo?, Geert-Jan Houben??
Technische Universiteit Eindhoven
PO Box 513, NL 5600 MB, Eindhoven, The Netherlands
Abstract. iFanzy is a personalized TV guide application aiming at of-
fering users television content in a personalized and context-sensitive
way. It consists of a client-server system with multiple clients and de-
vices such that the user can ubiquitously use TV set-top box, mobile
phone and Web-based applications to select and receive personalised TV
content. TV content and background data from various heterogeneous
sources is integrated to provide a transparent knowledge structure, which
allows the user to navigate and browse the vast content sets nowadays
available. Semantic Web techniques are applied for enriching and align-
ing Web data and (live) broadcast content. The resulting RDF/OWL
knowledge structure is the basis for iFanzy’s main functionality, like se-
mantic search of the broadcast content and execution of context-sensitive
Current Web-based applications are characterized by the fact that users use
many different types of devices to access content. As a consequence, engineering
such applications has to respect the different environments and capabilities to
ensure that the application adapts to the circumstances. A big advantage for the
purpose of personalization is that the user spends more time with these devices
than with the PC alone which gives more information to assess and model the
user’s situation. In the TV broadcasting domain the combination of multiple
devices such as mobile phone and TV with multiple applications accessible via
the Web shows how access to content will evolve .
At the same time we see the integration of background information that re-
lates to TV content which can help the user in selecting and using this data.
Much of this information is available via the Web, which leads to the signif-
icant trend that television viewers use their PC for the larger part of the TV
content access process. While advantageous this trend also implies an informa-
tion overload that cries out for adaptation to the user’s knowledge, preferences
and situation. Personalization in this setting can benefit from several different
?also affiliated with Vrije Universiteit, Amsterdam, the Netherlands
??also affiliated with Vrije Universiteit Brussel, Brussels, Belgium
kinds of integration: friends or relatives watching TV content together, inte-
grating (background) data from different connected applications and integrating
temporal and spatial-specific viewpoints in context modeling.
In this paper we present iFanzy - a personalized TV guide application aiming
at offering users television content in a personalized and context-sensitive way
(developed in collaboration with Stoneroos Interactive TV, Ltd.1). It is currently
available as a Web application and a TV set-top box front-end, while a mobile
version is under development. In a client-server model, iFanzy acts as a client
that uses a server framework called SenSee  as underlying data source. This
takes care of content integration, user modeling and content recommendation.
Several other TV recommender systems exist, e.g. AVATAR  which has
a focus on reasoning over TV content metadata and user preferences. iFanzy
differs from AVATAR mainly because of our focus on combining and integrat-
ing the information from several large and live datasources. Many systems exist
that focus on the recommendation part, for example the movie recommenda-
tion application MovieLens2that uses collaborative filtering. For an overview of
different recommendation strategies e.g. refer to .
2 Semantics-based Content Integration
The use of semantics is an important instrument in order to combine and inte-
grate the content from different applications and in this way to enhance person-
alization. In this sense iFanzy represents a large class of multi-device applications
with a high degree of interactivity where semantics is key to effective integration
. In our work we have applied a general strategy that supports this large class
of semantics-based applications, illustrated here in terms of iFanzy.
Step 1: Making TV metadata available in RDF/OWL
As a first step we make the relevant metadata from various data sources available
in RDF/OWL. In the current iFanzy demonstrator we use three live data sources,
online TV guides in XMLTV format (e.g. 1.2M triples for the daily updated
programs), online movie databases such as IMDB in custom XML format (e.g.
8M triples for 12K movies and trailers from Videodetective.com), and broad-
cast metadata available from BBC-backstage in TV-Anytime (http://www.tv-
anytime.org/) format (e.g. 92K triples, daily updated). Next to the live data we
also use the W3C OWL Time Ontology3to represent time information.
Step 2: Making relevant vocabularies available in RDF/OWL
Having the metadata available, it is also necessary to make relevant vocabularies
available in RDF/OWL. In iFanzy we did this in a SKOS-based manner for the
genre vocabularies (resulting in 5K triples), and for the TV-Anytime Genres,
the XMLTV Genres, and the IMDB Genres. All these genres play a role in the
classification of the TV content and the user’s likings (supporting the recommen-
dation). We also used WordNet 2.0 (http://www.w3.org/2006/03/wn/wn20/) as
published by W3C (2M triples) and the locations used in IMDB (60K triples).
Step 3: Aligning and enriching vocabularies/metadata
Here we did (1) alignment of Genre vocabularies, (2) semantic enrichment of the
Genre vocabulary in TV-Anytime, and (3) semantic enrichment of TV metadata
with IMDB movie metadata.
First, aligning the Genre vocabularies was a small semi-automated exercise
in which several translations were specified towards the TV-Anytime vocabulary,
such as the associations between xmltv:documentaire and tva:documentary,
between imdb:thriller and tva:thriller, and between imdb:sci-fi and
Second, for the semantic enrichment of the Genre vocabulary,
– based on the original XML Term hierarchy, skos:narrower relations are in-
troduced, for example between tva:news and tva:sport news.
– based on partial label matching, skos:related relations are defined, for
example between tva:sport news and tva:sport.
– background design knowledge has been the motivation for distinguishing
skos:related relations between siblings, such as between tva:rugby and
Third, in terms of semantic enrichment of the TV metadata (that can come
from different grabbers in different languages) we use from IMDB the country
AKA-titles to link each grabbed program to the associated concept in IMDB.
Step 4: Using the resulting RDF/OWL graph for recommendations
To recommend TV programs or movies, the resulting RDF/OWL graph is
extended with the user model in a format such that the eventual RDF/OWL
knowledge structure can be directly used for the recommendation. What happens
is that when user rates a program P, implicitly program P is rated as well as all
programs which are related in the knowledge structure. Moreover all programs
with a genre that is related to a genre of P are rated, as well as the genres
themselves via skos:related and skos:narrower relations. In imdb:persons
all actors, directors and persons associated with P are rated. In this way, ratings
are added to the user model, within the user’s context.
3 iFanzy Architecture
An important requirement for iFanzy is to provide this service in a ubiquitous
and responsive way, e.g. independent of the platform used or the current location
of the user. Therefore we opted for a client-server architecture, where the user
uses the iFanzy front-end with different devices connected to the SenSee server.
All heavy computation work is done at the server side. This ensures that virtually
any machine (including mobiles and set-top boxes) that can connect through the
Internet can be linked to the system.
The server deals with very large data collections of browsable content - hun-
dreds of thousands of programs from various sources, as well as knowledge struc-
tures used for recommendation and semantic search. Thus, SenSee should handle
the concurrent use of hundreds of potential users per server. Although we see Download full-text
many data-intensive Semantic Web applications, scalability is still an important
research issue for truly real-time Web-applications. In order to reach the desired
scalability we performed many optimization steps .
The recommendation part depends heavily on the quality of the system’s
knowledge of the user. To cope with the cold start, we devised a statistical
recommendation algorithm to find the most relevant programs based on a basic
set of user registration data. Further, iFanzy’s training algorithm allows to refine
the user data from the user’s behaviour and explicit feedback. The Web client,
for instance, tracks the clicks made on specific content items and the search
terms used. The set-top box on the other hand, monitors and stores the viewing
behaviour. The user can also utter specific likings (explicit feedback) to inform
the system what he/she appreciates.
4Conclusion and future work
Different versions of the different clients and server systems have been imple-
mented and successfully evaluated in collaboration with our commercial partner
Stoneroos. As future work we are redesigning the iFanzy frontend and SenSee
backend, based on our practical experiences, and we plan a next performance
optimization step with parallel query evaluation and load-balancing strategies.
Currently, an evaluation trial with 500 set-top boxes in Dutch households is
prepared together with Stoneroos.
1. Bjorkman, M., Aroyo, L., Bellekens, P., Dekker, T., Loef, E., Pulles, R.: Personalised
home media centre using semantically enriched tv-anytime content. In: EuroITV
2006 Conference. (2006) 156–165
2. Aroyo, L., Bellekens, P., Bjorkman, M., Houben, G.J.: Semantic-based framework
for personalised ambient media. Multimedia Tools and Applications 36(1-2) (2008)
3. Bellekens, P., van der Sluijs, K., Aroyo, L., Houben, G.J.: Engineering semantic-
based interactive multi-device web applications. In: Proceedings of the 7th Interna-
tional Conference on Web Engineering (ICWE’07). Volume 4607 of Lecture Notes
in Computer Science., Como, Italy, Springer (2007) 328–342
4. Blanco Fernndez, Y., Pazos Arias, J.J., Gil Solla, A., Ramos Cabrer, M., Lpez Nores,
M.: Bringing together content-based methods, collaborative filtering and semantic
inference to improve personalized tv. 4th European Conference on Interactive Tele-
vision (EuroITV 2006), (May 2006)
5. van Setten, M.: Supporting people in finding information: Hybrid recommender
systems and goal-based structuring. Telematica Instituut Fundamental Research
Series, No.016 (TI/FRS/016). Universal Press. (2005)
6. Bellekens, P., van der Sluijs, K., van Woensel, W., Casteleyn, S., Houben, G.J.:
Achieving efficient access to large integrated sets of semantic data in web appli-
cations. In: Proceedings of the 8th International Conference on Web Engineering
(ICWE’08), to be published