AVATAR: A Flexible Approach to Improve the

Personalized TV by Semantic Inference

⋆

Yolanda Blanco Fernández, Jose J. Pazos Arias, Alberto Gil Solla and

Manuel Ramos Cabrer

Department of Telematic Engineering, University of Vigo, 36310, Spain

Abstract. Both the TV recommender systems and search engines developed in

the Internet are intended to lighten the user burden, by offering them automat-

ically the required information, personalized according to their preferences or

needs. In last years, with the goal of improving these search engines, an important

research line has been developed in the context of the WWW, known as the Se-

mantic Web. The Semantic Web describes the resources by metadata and reasons

on them by discovering new knowledge. Taking the advantage of the Semantic

Web in the ﬁeld of the personalized TV, we propose an intelligent assistant named

AVATAR, which uses the semantic inference as a novel recommendation strategy.

This approach allows to overcome an important limitation identiﬁed in the per-

sonalization strategies adopted in other systems: an excessive similarity between

the programs known by the user and those suggested by the recommender. In

this regard, our approach diversiﬁes and personalizes the elaborated recommen-

dations, by inferring semantic associations of different nature between the user

preferences and the suggested TV contents. This inference process requires a for-

mal representation both the knowledge of our application domain, and the user

preferences. In this regard, we resort to an OWL ontology to identify resources

and relations typical in the TV ﬁeld, and to reason about them.

1 Introduction

The adoption of different standards for Digital TV (DTV) envisages a scenario in which

the users can access to a greater number of audiovisual contents, transmitted together

with interactive applications. This situation causes the viewers feel disoriented among a

massive amount of irrelevant information. To address effectively this problem it is nec-

essary to resort to tools —named TV recommender systems [1]— which lighten the user

burden by offering automatically programs personalized according to their preferences.

The functionality of these recommenders is very similar to the goal pursued by the

search engines developed in the Web, whose utility is undeniable due to the overwhelm-

ing information available for their users. In last years, the so-called Semantic Webhas

become a relevant research line. The Semantic Web uses metadata for describing the

Web resources, so machines can understand these descriptions and infer semantic rela-

tionships between them. Consequently, the search engines for the Semantic Web over-

come clearly the simple syntactic approach adopted in tools like Google.

⋆

Work supported by the Ministerio de Educación y Ciencia Research Project TSI2004-03677

Blanco Fernández Y., J. Pazos Arias J., Gil Solla A. and Ramos Cabrer M. (2005).

AVATAR: A Flexible Approach to Improve the Personalized TV by Semantic Inference.

In Proceedings of the 1st International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces, pages 69-78

DOI: 10.5220/0001421100690078

 SciTePress

Taking the advantage of the Semantic Web, the TV recommenders can also be

improved by including capabilities of semantic inference. This way, here we present

AVATAR (AdVAnced Telematic search of Audiovisual contents by semantic Reasoning),

a TV recommender which suggests to a user contents semantically related to those

he/she watched in the past. This novel recommendation strategy enhances the offered

suggestions by discovering appealing relationships, unknown in previous approaches.

Such relationships —named semantic associations— are inferred from the knowledge

learned by our intelligent assistant.

Such inference process permits to overcome a drawback identiﬁed in previous rec-

ommendation approaches, that is, the excessive similarity between the contents known

by the user and those suggested by the system. To diversify the elaborated sugges-

tions, our proposal considers a complete classiﬁcation of semantic associations of di-

verse nature. In this regard, the kind of discovered association and the diversiﬁcation

are closely related in our approach: the more direct the inferred relationships between

the user preferences and the suggested programs, the less diversiﬁed this recommenda-

tion. Conversely, the discovery of indirect associations —not supported in the existing

approaches— allow to suggest varied TV contents. It is worth noting that the diversiﬁ-

cation introduced in our system by the semantic inference is especially useful in the TV

domain, where neither viewing habits of the users nor their preferences follow a homo-

geneous pattern. Consequently, the different types of associations permit to adjust the

offered recommendation to each viewer, leading to successful personalized suggestions.

To carry out such inference process our system requires a mechanism to represent

formally the knowledge of our speciﬁc application domain. In the Semantic Web, one of

the well-known methodologies intended for this purpose are the ontologies. So, we have

implemented an ontology —using the OWL (Web Ontology Language) language—

where resources and relations typical in the TV domain (categories of programs, genres,

credits involved in them, etc.) have been identiﬁed by means of classes, properties and

speciﬁc instances. This way, our approach focuses on the explicit properties deﬁned in

the TV ontology, with the goal of inferring implicit semantic associations between the

instances stored in this knowledge base.

On the other hand, taking into account that the ontologies are intended to reuse

efﬁciently the knowledge represented in them [2], our approach has also used the TV

ontology to model the user preferences, deﬁned in personal proﬁles.

This paper is organized as follows: Sect. 2 reviews the related work and highlights

the main differences with the aims of our proposal. Sect. 3 presents the structure of

our OWL ontology and deﬁnes the user proﬁles handled in AVATAR. Sect. 4 describes

the recommendation strategy used in our tool, including a complete classiﬁcation of the

supported semantic associations and explanatory examples. Finally, Sect. 5 summarizes

the main conclusions and discusses future work.

2 Related Work

The personalization of services, according to user preferences, can be located in the

80’s and bound to Internet. Personalization intends to resolve the problems caused by

the more and more available information, ﬁrstly in News groups and later in the WWW.

It is in this context where the three main categories of systems for contents personaliza-

tion are identiﬁed. From the perspective of the information which is analyzed to elab-

orate recommendations, the approaches can be classiﬁed in: (i) content-based methods

which compute personalized recommendations by comparing content representations

of previously liked items, with descriptions of items still unknown to the user; (ii) col-

laborative methods, which intend to recommend contents which have been interesting

for users with similar preferences; and (iii) hybrid methods, which combine both meth-

ods to meet their different advantages [3]. Some well-known recommenders are the

content-based NewsWeeder [4] and WebWatcher [5]; the collaborative Phoaks [6] and

GroupLens [7]; and the hybrid Fab [8].

We have explored personalization systems for programming guides in the TV do-

main. As in the Internet context, the ﬁnal aim is to avoid users managing such a huge

amount of information. The ﬁrst ones of these systems were offered from Web sites, for

instance PTV [9]. Users register themselves in the PTV server (Web site) and then they

can access to personalized programming guides presented as HTML or WML pages.

The system incorporates user proﬁles, content-based reasoning and collaborative meth-

ods to make recommendations. When registering a new user, the system creates a proﬁle

which stores preferences about programs, channels, genres, timetables, etc.

These systems are able to provide a good service, as it is the case of PTVPlus,

successor of PTV. Today PTVPlus is working in UK and Ireland and it provides per-

sonalization services of about a hundred of TV channels to thousands of people. How-

ever, two main lacks can be identiﬁed: the way in which user information is gathered

to elaborate the proﬁle, and the inability to maintain historic logs about what programs

have been watched by what users. So, in these systems, it is not possible to deploy

implicit techniques for managing the above information. The only solution is to use

an explicit technique which elaborates an initial proﬁle –from the information entered

at registration– and after that, the feedback information (watched programs, changes in

preferences, etc) needed for the explicit interaction and collaboration of the user through

speciﬁc Web pages. Despite the fact that good recommendations can be achieved by this

explicit mechanism, it entails the user to do a lot of work. Unfortunately, the user is not

often willing to ﬁll up a tedious form about his preferences, and then to access again

and again to the WWW server in order to mark the programs.

Several ingenious solutions have been proposed to solve the problem of storing

the user viewing history, such as the GuideRemote system. This proposal is based on

a universal remote control that, after being connected to a PC, is able to download

from the appropriate website the personalized programming guide for the following

seven days, being also able to maintain information about the TV programs previously

watched by the user and, after that, to inform the server. However, a personalization

system running in the user receiver should be the most logic solution because all the

information about the user and TV interaction is always available.

Having this information in mind, in this kind of systems is possible to set up implicit

learning techniques about user interests and behavior. Additionally, this information can

be continuously improved by taking into account all decisions adopted by the user any

time he interacts with the system. This allows a great reduction in the information that

must be explicitly given by the user so, in this case, the ﬁrst time the user turns on the

system only must inform about a few characteristics to built a preliminary proﬁle.

Therefore, it is possible to ﬁnd recommender systems [10] based on a multiagent

architecture running in the user receiver. In these proposals specialized agents collab-

orate to obtain and combine the information broadcast with the information available

in the Internet about TV programming guides. Besides that, they are also able to learn

the user preferences and behavior in an implicit way, which allows having better rec-

ommendations in future. Likewise, [11] uses three different sources to build the user

proﬁle: his implicit viewing history, his explicit preferences and his interaction with the

system. In this kind of systems, Bayesian techniques or those based on decision trees

are the most used to allow the implicit learning of proﬁles.

Our approach differs from these works, given that our proposal uses several mech-

anisms to represent the knowledge about the TV domain, taking the advantage of pre-

vious experience in the Semantic Web. In order to carry out inference processes, our

system requires generic descriptions about TV programs, that is, metadata about these

TV contents.

Our proposal is based on ﬁnding out semantic relations between TV programs, by

reasoning about the content descriptions provided by their metadata. The issue of the

inference of complex relations has been only addressed out of the TV domain. Specif-

ically, [12] deﬁnes this kind of associations for national security applications over a

RDF(S) set. Its goal is to discover relationships between two instances speciﬁed by

the user. Our approach differs from it because AVATAR stems from the viewer pref-

erences, and ﬁnds out which contents are semantically related to them in a signiﬁcant

way. In other words, [12] resorts to two instances and AVATAR considers only one and

discovers the second one, that is, the set of suggested programs.

3 Background

3.1 The TV Ontology

In the Semantic Web, ontologies are well-known methodologies for representing the

knowledge of a speciﬁc application domain. In our proposal we have also resorted to

an ontology, implemented using the OWL language, to conceptualize the TV domain.

This ontology, referred to as O, consists of two parts: (i) A knowledge base (KB

)

which contains classes and properties hierarchically organized, and (ii) a description

base (DB

) containing a set of instances of these classes related by the properties

contained in KB

. These instances are TV programs of different categories, actors,

directors involved in them, genres, etc.

To identify relations among different resources, the concept of property sequences [12]

is used. A property sequence P S is deﬁned as a ﬁnite set of properties [P

, P

, . . ., P

which joins different classes stored in the knowledge base of the ontology O. This way,

the length of a sequence is deﬁned as the number of properties contained in it.

Given that this ontology contains speciﬁc instances of each class, it is also possible

to deﬁne a property sequence instance, represented by lower case letters. So, ps is a

chain of properties which join different instances present in DB

. The ﬁrst node of

Formula1

Broadcasting

Terror

Movies

Actors

Sport

Places

Formula1

Drivers

Soccer

Broadcasting

Action

Movies

Comedy

Movies

Sport

Documentaries

Soccer

Players

Sporting

Events

C. Eastwood

Stadium Amsterdam Arena

Owen

F. Alonso

producerOf

hasPlace

hasPlayer

CollaboratesIn

hasPlace

hasDriver

Soccer

Players

D. Beckham

hasStarringActor

hasSupportingActor

hasParticipant

hasName

Soap

Operas

Quiz Shows

Actresses

presenterOf

hasSupporting

Actress

hasStarring

Sport

Places

Stadium Yokohama Julia Roberts

Actress

hasNamehasName

hasParticipant CollaboratesIn

CollaboratesIn

Fig.1. Excerpt of DB

the instance ps is named origin, and the last one terminus. Besides, two functions have

been used to return the classes contained in the sequence P S —PS.NodesOfPS()—

and the instances contained in ps —ps.PSNodesSequence().

Just as its name suggests, a property sequence P S (and an instance ps) fulﬁlls that

the class (instance) which appears as range in the property P

, is the same one which

ﬁgures as domain in P

i+1

. To clarify the above concepts, an excerpt of our ontology is

shown in Fig. 1, where the circles represent speciﬁc instances of classes with the same

name, the arrows are properties, and the rectangles are concrete values of data types.

Here we can see different property sequences. For example, in the left side of the

ﬁgure, the sequence ps = [hasStarringActor, producerOf, hasPlace, hasName] is shown.

This property sequence allows to identify an action movie in which Clint Eastwood is

involved as starring actor. Besides, this actor is also producer in a terror movie happened

in the Stadium Amsterdam Area. This way, the nodes of this sequences are the follow-

ing instances: ps.PSNodesSequence() = [Action Movies, Actors, Terror Movies,

Sport Places], where Action Movies is the origin and Sport Places is the terminus of the

sequence. On the other hand, the nodes contained in the set PS.NodesOfPS()would

be the classes to which these instances belong. Note that in Fig. 1, we are representing

instances and therefore, here Action Movies is a concrete instance of the class identiﬁed

with the same name.

Next, we show how the knowledge represented in O is efﬁciently reused to model

the user preferences.

3.2 The User Proﬁles

Each user proﬁle deﬁned in AVATAR stores personal information about the user and

his/her preferences. To model these preferences we have used a dynamic subset of our

OWL ontology, built incrementally by adding new classes, properties and instances —

extracted from O — as the system knows additional information about his/her interests.

For that reason, these proﬁles are named in AVATAR ontology-proﬁles. Speciﬁcally,

when AVATAR knows a new content, the system adds to the proﬁle: (i) this instance,

(ii) the class referred to it, (iii) the hierarchy of superclasses deﬁned in KB

related to

this class and ﬁnally, (iv) some properties deﬁning the main characteristics of this TV

content.

For example, let us suppose that a user watched a Formula 1 race in which the driver

Fernando Alonso is involved, and the action movie starring Clint Eastwood mentioned

in the previous section. In this case, AVATAR adds to the user proﬁle the instances Ac-

tion Movies and Formula1 Broadcasting shown in Fig. 1, the classes to which these

instances belong (they are classes with the same name than the mentioned instances),

the superclasses Sport Programs and Broadcasting Programs, with respect to Formula1

Broadcasting, and the superclasses Movies with respect to Action Movies. Finally, some

properties are also included in the user ontology-proﬁle, such as hasStarringActor, has-

Driver, etc, together with their respective values.

It is worth noting that the ontology-proﬁles handled in AVATAR improve greatly

the plain lists adopted in previous approaches. These lists allows to deﬁne the user

preferences, but not organize them in a structural and hierarchical way, causing that

discovery of new knowledge is clearly hampered.

On the other hand, to maintain the user preferences permanently updated, our pro-

posal deﬁnes three indexes related to each class and each instance contained in his/her

proﬁle. These indexes are updated in an automatic way from the actions carried out by

the TV viewer as we describe in [13]. They are the DOI (Degree Of Interest), the Conﬁ-

dence and the Relevance indexes. The former one measures the level of interest referred

to each class/instance; the second one quantiﬁes the success or failure of the system in

the previous recommendations; and, ﬁnally, the Relevance index —which combines the

previous ones— is used to rank the suggested programs. Finally, note that user proﬁles

store both those contents the user likes (called positive preferences) and those he/she

dislikes (negative preferences).

4 A Recommendation Strategy based on Semantic Inference

Once formalized the TV ontology and the user proﬁles, we describe the cornerstone of

our recommendation process: the semantic associations supported in our approach.

4.1 The Semantic Associations

The semantic associations included in our approach are built taking as a starting point

the relations established between the properties, and between the property sequences

deﬁned in O. First, we formalize the information used to deﬁne these relations.

Let P S

= [P

, . . . , P

] and P S

= [Q

, . . . , Q

] be two property sequences,

and ps

two instances of them, respectively.

Our approach deﬁnes a relation between any pair of properties as follows:

- Semantic Nexus between Properties (↔

): P and Q are related by a semantic

nexus —denoted as P ↔

Q— if either of the conditions is satisﬁed: (i) they the same

property, (ii) P is superproperty of Q, (iii) Q is superproperty of P or (iv) P and Q are

sibling properties.

Example 1: In Fig. 1 we can establish several semantic nexus between properties:

for instance, producerOf ↔

presentedOf is true because both properties are sibling.

Analogously, hasSupportingActress ↔

hasStarringActor is also veriﬁed.

Regarding the property sequences, two relations between them can be deﬁned:

- ρ - Isomorphic Property Sequences (

∼

). P S

and P S

are ρ−isomorphic —

denoted as P S

∼

P S

— if all their properties are related by semantic nexus one-to-

one, that is: ∀ i, 0 ≤ i ≤ m, P

↔

is veriﬁed.

Example 2: Let us consider the sequences shown below. For clarity, the classes are

represented in italics, and the properties in brackets.

P S

: Soap Operas [hasSupportingActress] Actresses [presenterOf] Quiz Shows

P S

: Action Movies [hasStarringActor] Actors [producerOf] Terror Movies

Taking into account the semantic nexus shown in Example 1, it is not hard to see

that the above sequences are ρ-isomorphic.

- Joined Property Sequences (⊲⊳

). P S

and P S

are joined —P S

⊲⊳

P S

—

if they contain at least a common class. In other words, P S

⊲⊳

P S

is veriﬁed if

P S

.NodesOfPS() ∩ P S

.NodesOfPS() is a nonempty set. This way, a join class C (i.e.

C ∈ P S

.NodesOfPS() ∩ P S

.NodesOfPS()) is named join node.

Example 3: The sequence properties P S

: Terror Movies [hasPlace] Sport Places

and P S

: Sporting Events [hasPlace] Sport Places, share the join node Sport Places,

and consequently they are ρ-joined.

Taking into account the aforementioned relations, [12] deﬁnes signiﬁcant semantic

associations between two entities deﬁned in a RDF(S) set. Some of them are especially

useful in our recommendation process, and, for that reason, they have been assumed in

our approach. A brief review of such associations is presented here:

I. ρ - pathAssociated: ρ - pathAssociated (x,y) is true if there exists a property

sequence instance ps and, either x and y are the origin and terminus of ps respectively,

or vice versa, i.e. y is origin and x is terminus.

Example 4: We can illustrate this kind of semantic association by resorting to any of

the property sequences shown in previous examples. So, we can deﬁne the sequence ps

by extracting speciﬁc instances of the classes shown in P S

in Example 2. This way, it is

possible to deﬁne an association ρ-pathAssociated between the instance corresponding

to the soap opera represented in Fig. 1, and the instance referred to the quiz show.

II. ρ - joinAssociated: Let P S

and P S

be two joined property sequences with a

join node C (i.e. P S

⊲⊳

P S

). Besides, assume that there exist ps

and ps

which

contain two instances —equal or different— belonging to this class C.

ρ - joinAssociated (x,y) is true if (i) x is the origin of ps

and y is the origin of ps

or (ii) x is the terminus of ps

and y is the terminus of ps

Example 5: Considering the joined property sequences P S

and P S

deﬁned in

Example 3, it is possible to deﬁne two sequences ps

and ps

by extracting speciﬁc

instances of the classes Terror Movies, Sporting Events and Sport Places. So, an associ-

ation of the form ρ-joinAssociated is established between the terror movie represented

in Fig. 1, and the sporting event. Note that the join node is the Sport Places class, how-

ever ps

y ps

contain different instances of this class: ps

is related to the stadium

Amsterdam Arena and, on the other hand, ps

is referred to the stadium Yokohama.

III. ρ - isoAssociated. ρ - isoAssociated (x,y) is true if there exist two ρ-isomorphic

property sequences P S

and P S

(P S

∼

P S

), and there exist ps

and ps

whose

origins are x and y, respectively.

Example 6: This example is easy to extract from the isomorphic property sequences

P S

and P S

shown in Example 2. So, an association of the form ρ-isoAssociated is

established between speciﬁc instances corresponding to the soap opera starring Julia

Roberts, and the action movie where Clint Eastwood is involved (in Fig. 1, ρ - isoAsso-

ciated (Soap Operas, Action Movies)).

With the goal of enhancing the inferential processes in AVATAR, several semantic

associations not considered in [12], are deﬁned in our proposal. Taking into account the

necessary balance between the personalization and the diversiﬁcation of recommenda-

tions, we propose the so-called mixed semantic associations, by taking as a starting

point the ρ-pathAssociated association, given that it is the most direct and meaningful

relationship out of the ones described in [12]. These mixed associations are shown next.

IV. ρ - join-pathAssociated. Just as its name suggests, this association mixes the

following ones: ρ - pathAssociated and ρ - joinAssociated. So, ρ - join-pathAssociated

(x,z) is true if either of the conditions is satisﬁed:

( ρ - pathAssociated (x,y) and ρ - joinAssociated (y,z) ) or

( ρ - joinAssociated (x,y) and ρ - pathAssociated (y,z) )

Example 7: In Fig. 1, the associations ρ-pathAssociated (Formula 1 Broadcasting,

Sport Documentaries) and ρ-joinAssociated (Sport Documentaries, Soccer Broadcast-

ing) can be identiﬁed. Consequently, the association ρ-join-pathAssociated (Formula 1

Broadcasting, Soccer Broadcasting) is direct.

V. ρ - cp-pathAssociated. Similarly, ρ - cp-pathAssociated (x,z) is true if:

( ρ - pathAssociated (x,y) and ρ - cpAssociated (y,z) ) or

( ρ - cpAssociated (x,y) and ρ - pathAssociated (y,z) )

VI. ρ - iso-pathAssociated. Finally, ρ - iso-pathAssociated (x,z) is satisﬁed if:

( ρ - pathAssociated (x,y) and ρ - isoAssociated (y,z) ) or

( ρ - isoAssociated (x,y) and ρ - pathAssociated (y,z) )

4.2 The Process of Recommendation

The open and modular architecture proposed for AVATAR [14] allows to add new rec-

ommendation strategies without including signiﬁcant modiﬁcations. For its novelty in

the personalization domain, here we focus on the strategy based on semantic inference

proposed in this paper. Regarding this, two phases can be distinguished:

- Phase of Semantic Inference: First, AVATAR explores the property sequences

deﬁned in DB

whose origin are instances contained in the user ontology-proﬁle. The

associations supported in our proposal are easy to infer by stemming from these se-

quences. For that purpose, it is only necessary to check if the conditions shown in

Sect. 4.1 for each kind of association, are veriﬁed by these sequences. For example,

two sequences which share a common class originate an association of the form ρ-

joinAssociated between their respective origins (or terminus).

It is worth noting that in order to avoid retrieving a massive amount of meaningless

associations, our approach resorts to a ﬁltering methodology previous to the aforemen-

tioned inference. This methodology ignores those instances of the analyzed sequences

which are not relevant enough according to the personalization requirements, and re-

trieves only signiﬁcant associations. Obviously, the quantiﬁcation of such relevance

depends on the user preferences: if an instance is closely related to the programs the

user liked, this one is not ﬁltered and consequently, it will be contained in the inferred

semantic associations. This phase returns a set of programs semantically related to the

user preferences, together with the type of discovered association.

- Phase of Ranking: This list of programs must be ranked and shown to the users.

In this process AVATAR considers two factors. The ﬁrst one is the index Relevance of

each instance in the user proﬁle (remember Sect. 3.2). Note that if a program is very

appealing for the user, its relevance index takes high values. This way, those inferred

programs which are related to the user preferences with highest relevance, will be rank

in provisional top positions of the suggestion. To conﬁrm this position in the recommen-

dation, AVATAR resorts to some parameters relevant in a personalized TV environment,

such as the age rating of the program, the viewer favorite time for watching TV, etc.

5 Conclusions and Further Work

In this paper we have proposed AVATAR, a recommender system which takes the ad-

vantage of the Web Semantic and extends it to the domain of personalized TV. Its main

contribution is a reasoning process about the user preferences and descriptions of TV

contents. The recommendation strategy applied in our tool is based on suggesting a

user TV programs which are semantically related to those watched in the past. Our

approach differs from other existing proposals because semantic inference capabilities

have never been considered in a personalization environment. Such inference process

allows AVATAR to adjust in a ﬂexible way to the kind of suggestion required by each

viewer: if the user wants to watch programs similar to ones he liked in the past, AVATAR

infers very direct associations between the suggested programs and his/her preferences.

On the contrary, if the user does not mind watching varied programs, less similar to

those he knows but always related to them, our inference methodology considers in-

direct associations. So, AVATAR discovers contents especially appealing for the user

which would not be suggested by previous approaches.

It is worth noting that our inference methodology is not only valid for the TV do-

main. As mentioned in Sect. 4.2, the phase of inference reasons about a generic OWL

ontology and the user preferences, by discovering semantic associations meaningful ac-

cording to his personal interests. On the contrary, those parameters speciﬁc of the TV

domain are considered in a separate phase (the phase of ranking). Such decomposition

favors the reuse of our inference approach in other personalization applications.

For example, our proposal could be applied in recommender systems for the Se-

mantic Web, where few proposals have been deﬁned up to now. So, a possible example

for motivating the usefulness of our approach in the Semantic Web, could be a system

which suggests to a user travel destinations, by considering relations between coun-

tries according to their customs, culture, etc. For that purpose, it is only necessary to

conceptualize this new domain by an OWL ontology, and to know the user preferences.

At this moment, we have a ﬁrst prototype of AVATAR by which appealing recom-

mendations have been obtained. Our goal is to extend these encouraging results to a real

scenario with a greater number of users. For that reason, we are working on statistical

studies by considering different population groups with diversiﬁed characteristics, such

as genre, age, cultural level, etc. This way, we evaluate the usefulness of the semantic

inference as novel recommendation strategy both in the TV domain and out of it.

References

1. Ardissono L., Kobsa A. and Maybury M. In: Personalized Digital TV: Targeting Programs

to Individual Viewers. Kluwer Academic Publishers (2004)

2. Davies J., Fensel D. and van Harmelen F. In: Towards the Semantic Web: Ontology-driven

Knowledge Management. John Wiley & Sons (2003)

3. Burke R.: Hybrid Recommender Systems: Survey and Experiments. User Modeling and

User-Adapted Interaction 12 (2002) 331–370

4. Lang K.: Newsweeder: Learning to ﬁlter netnews. In: Proc. of 12th International Conference

on Machine Learning. (1995)

5. Armstrong R., Freitag D., Joachims T. and Mitchell T.: WebWatcher: A Learning Apprentice

for the World Wide Web. In: AAAI Spring Symposium on Information Gathering. (1995)

6–12

6. Terveen L., Hill W., Amento B., McDonald D. and Creter J.: PHOAKS: A system for sharing

recommendations. Communications of the ACM 40 (1997) 59–62

7. Konstan J., Miller B., Maltz D., Herlocker J., Gordon L. and Riedl J.: GroupLens: Applying

collaborative ﬁltering to Usenet news. Communications of the ACM 40 (1997) 77–87

8. Balabanovic M. and Shoham Y.: FAB: Content-Based Collaborative Recommender. Com-

munications of the ACM 40 (1997) 66–72

9. Smyth B. and Cotter P.: A personalized television listings service. Communications of the

ACM 43 (2000) 107–111

10. Diﬁno A., Negro B. and Chiarotto A.: A multi-agent system for a personalized electronic

program guide. In: Proc. of the Workshop on Personalization in Future TV (TV-02). (2002)

11. Kurapati K., Gutta S., Schaffer D., Martino J. and Zimmerman J.: A multi-agent tv recom-

mender. In: Proc. of the Workshop on Personalization in Future TV (TV-01). (2001)

12. Anyanwu K. and Sheth A.: ρ-Queries: Enabling Querying for Semantic Associations on the

Semantic Web. In: Proc. of the 12th International World Wide Web Conference (WWW-03),

ACM Press, 1-58113-680-3/03/0005 (2003) 690–699

13. Blanco Y., Pazos J., Gil A., Ramos M., López M. and Barragáns B.: AVATAR: Modeling

Users by Dynamic Ontologies in a TV Recommender System based on Semantic Reasoning.

In: Proc. of the 3rd European Conference on Interactive Television (EuroITV-05). (2005)

14. Blanco Y., Pazos J., Gil A., Ramos M., García J., Díaz R., Fernández A., Barragáns B. and

López M.: AVATAR: Advanced Telematic Search of Audiovisual Contents by Semantic

Reasoning. In: Proc. of the 4th Workshop on Personalization in Future TV (TV-04). (2004)