AVATAR: An Approach based on Semantic Reasoning to Recommend Personalized TV Programs Yolanda blanco Jose j pazos Alberto Gil Manuel ramos yolanda@det. uvigo es jose @uvigo es agil@det vigo. es mramos @uvigoes Ana fernandez Rebeca p diaz Martin Lopez Belen Barragans avilas@det. uvigoes rebeca @det. uvigoes mInores @det. uvigo es belen@det. uvigoes Department of Telematic Engineering University of Vigo, 36310, Spain ABSTRACT could feel disoriented as they look for interesting programs among In this paper a TV recommender system called AVATAR(AdVAnce massive amounts of irrelevant information. For that reason, sys- Telematic search of Audiovisual contents by semantic Reasoning) tems whose task is to recommend TV programs personalized ac is presented. This tool uses the experience gained in the field of corain to the user preferences, are clearly necessar In last years, the Tv recommender systems have become a rel- end users. The main contribution of our system is a process of se- evant issue for the research community, causing the appearance of -provided by means of metainformation-and on the viewer pref expert systems and neural networks, among others. However, these erences-contained in personal profiles. Such process allows to inference strategies have a common drawback related to the capa- diversify the offered suggestions maintaining the personalization, bility of reasoning about the Tv content semantics. given that the aim is to find contents appealing for the users, which se main contribution to the field of the TV personalization tools is Here the framework proposed for this reasoning is introduced, the use of the Semantic Web technologies [ I]. The goal is the devel- by including(i) the OWl ontology chosen to represent the knowl ment of an enhanced system by applying a reasoning process or edge of our application domain, (ii) the organization of the user the semantics of TV contens-provided by means 1o dIscover metadata- files, (iii) the query language LiKO, which is intended to browse and on the user preferences. Such reasoning is able the ontology and (iv) the semantic relations inferred from the sys- semantic relations among different programs. This way, the view ers are provided with diverse and personalized recommendations An overview of the architecture proposed for our prototype is de- Categories and Subject Descriptors scribed in [2], where the use of several knowledge inference meth- ods is shown, mainly Bayesian techniques and the aforementioned H3.3 [Information Storage and Retrieval]: Information Search emantic reasoning. For its relevance and novelty, in this paper we and Retrieval focus on this reasoning process This paper is organized as follows: Sect. 2 describes briefly General terms the framework of reasoning proposed in AVATAR. This framework consists of four elements: (i)an OwL ontology that represents the Languages, Theory, Management knowledge about the TV domain, (i) the user profiles, (ii)a query language for browsing the Tv ontology and (iv) the inference of Keywords semantic relations from the knowledge base of the system. Finally Semantic Web, TV recommender system, ontologies, inference of Sect. 3 reports some conclusions and discusses the future work. semantic relations 2. THE REASONING FRAMEWORK 1. INTRODUCTION The Digital Television allows the users access a large number of 2.1 The TV ontology TV programs and interact with applications transmitted along with To obtain a personalized TV recommender system based on se the traditional audiovisual contents. In this scenario. the viewers man ning, a mechanism for representing formally the knowl edge of our application domain is required. In the field of the search Project TSDy the Ministerio de Educacion y Ciencia Re- Semantic Web, the ontolos ax chnologies that represent and 04-03677 share this knowledge efficie TV domain by implementing an ontology according to the Owl language. This ontology allows to ident hierarchies of Copyright is held by the author/owner(s) classes and properties-entities related to the tv domain, suc -14, 2005, Chiba, Japan as different categories of programs, credits involved in them, etc ACM1-59593-051-5/05/00 1078
AVATAR: An Approach based on Semantic Reasoning to Recommend Personalized TV Programs∗ Yolanda Blanco yolanda@det.uvigo.es José J. Pazos jose@uvigo.es Alberto Gil agil@det.uvigo.es Manuel Ramos mramos@uvigo.es Ana Fernández avilas@det.uvigo.es Rebeca P. Díaz rebeca@det.uvigo.es Martín López mlnores@det.uvigo.es Belén Barragáns belen@det.uvigo.es Department of Telematic Engineering University of Vigo, 36310, Spain ABSTRACT In this paper a TV recommender system called AVATAR (AdVAnce Telematic search of Audiovisual contents by semantic Reasoning) is presented. This tool uses the experience gained in the field of the Semantic Web to personalize the TV programs shown to the end users. The main contribution of our system is a process of semantic reasoning carried out on the descriptions of the TV contents —provided by means of metainformation— and on the viewer preferences — contained in personal profiles. Such process allows to diversify the offered suggestions maintaining the personalization, given that the aim is to find contents appealing for the users, which are related semantically to their programs of interest. Here the framework proposed for this reasoning is introduced, by including (i) the OWL ontology chosen to represent the knowledge of our application domain, (ii) the organization of the user profiles, (iii) the query language LIKO, which is intended to browse the ontology and (iv) the semantic relations inferred from the system knowledge base. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval General Terms Languages, Theory, Management Keywords Semantic Web, TV recommender system, ontologies, inference of semantic relations 1. INTRODUCTION The Digital Television allows the users access a large number of TV programs and interact with applications transmitted along with the traditional audiovisual contents. In this scenario, the viewers ∗Work supported by the Ministerio de Educación y Ciencia Research Project TSI2004-03677 Copyright is held by the author/owner(s). WWW 2005,, May 10–14, 2005, Chiba, Japan. ACM 1-59593-051-5/05/0005. could feel disoriented as they look for interesting programs among massive amounts of irrelevant information. For that reason, systems whose task is to recommend TV programs personalized according to the user preferences, are clearly necessary. In last years, the TV recommender systems have become a relevant issue for the research community, causing the appearance of different approaches, such as Bayesian techniques, decision trees, expert systems and neural networks, among others. However, these inference strategies have a common drawback related to the capability of reasoning about the TV content semantics. So, our approach proposes a recommender, called AVATAR, whose main contribution to the field of the TV personalization tools is the use of the Semantic Web technologies [1]. The goal is the development of an enhanced system by applying a reasoning process on the semantics of TV contens —provided by means of metadata— and on the user preferences. Such reasoning is able to discover semantic relations among different programs. This way, the viewers are provided with diverse and personalized recommendations. An overview of the architecture proposed for our prototype is described in [2], where the use of several knowledge inference methods is shown, mainly Bayesian techniques and the aforementioned semantic reasoning. For its relevance and novelty, in this paper we focus on this reasoning process. This paper is organized as follows: Sect. 2 describes briefly the framework of reasoning proposed in AVATAR. This framework consists of four elements: (i) an OWL ontology that represents the knowledge about the TV domain, (ii) the user profiles, (iii) a query language for browsing the TV ontology and (iv) the inference of semantic relations from the knowledge base of the system. Finally Sect. 3 reports some conclusions and discusses the future work. 2. THE REASONING FRAMEWORK 2.1 The TV Ontology To obtain a personalized TV recommender system based on semantic reasoning, a mechanism for representing formally the knowledge of our application domain is required. In the field of the Semantic Web, the ontologies are technologies that represent and share this knowledge efficiently. So, we have conceptualized the TV domain by implementing an ontology according to the OWL language. This ontology allows to identify —by hierarchies of classes and properties— entities related to the TV domain, such as different categories of programs, credits involved in them, etc. 1078
Table 1: Query operators of the Liko language Operators 0 a class defined in the ontology. Direct subclasses related to th I a class defined in the ontology. Direct superclasses referred to the input class ed to the Instances of classes defined in the ontology. erties where the input class appears as domain. he ontolo To define specific information on which apply the precepts of not hard to see these contents are related, because the same perso our semantic reasoning approach, instances of the ontology classes is involved in both of them. have been included in this knowledge base. These instances are re- After the inference process, the goal is to rank the semantic path ferred to specific TV programs extracted from databases that store detected by extracting the most relevant ones. The most signifi- oundant information about TV scheduling, such as win imdb. com. cant relations involve the instances contained in the user ontology Regarding the properties, note they establish explicit relations among profile and other instances related to the former ones from a seman- the instances contained in the ontology. The reasoning process pr tic point of view. We refer to this new relation as semantic nexus posed in AVATAR takes as starting point these explicit relations and So, there exists a semantic nexus between two instances of classes infers other implicit ones from them. (or two properties) when these classes(properties) are siblings or are related by the subsumption principle(relations"is-a) 2.2 The User Profiles The relations defined here . which are easily discovered by the As we said in Sect. 2. 1, our system must reason on the LIKo query operators shown in Table 1, support the proposed preferences. For that reason, a mechanism for representin soning process and improve the personalized recommendations pre mally the user profiles is required. Our approach proposes t iting the knowledge contained in the TV ontology for that purpose So, we define the user profiles as OWL ontologies (which we 3. CONCLUSIONS AND FURTHER WORK ill refer to as ontology profiles), built incrementally as the syster In this paper a TV recommender system has been presented. The receives information about the user viewing behavior. This way, the main contribution of our tool is the application of a semantic rea ontology-profiles contain hierarchies of classes, specific instances and properties that are extracted from the TV ontology. In addition, soning process on the descriptions of the programs and the user the system also stores the level of interest for a user, referred to each preferences. This approach allows to infer hidden relations in the instance defined in his/her ontology-profile system knowledge base, taking as starting point the instances con- A user modeling based on ontologies allows avatar to com- tained in an Owl ontology about the tv domain, and the proper pare different profiles by applying the knowledge about the Tv do- ties that define explicit relations among them. The existing approaches do not reason about the semantics of main,beyond a simple syntactic comparison tackled in previous the offered programs. This way, the inference and reasoning pro- approaches. So, our recommender is able to find programs sema tically similar to those ones that have been of interest for the user cesses are greatly hampered. Our proposal provides the users with varied and personalized recommendations, by discovering sema 2.3 LIKO: A query language tic relations among the programs of interest for them and other ones AVATAR takes as starting point the user ontology-profille to offer him/her a perso alized recommendation. Once stances de- On the other hand, note that our proposal of semantic reasoning fined in this profile have been located in the TV ontology, it is nec- is not only valid for a Tv recommender system like AVATAR, but also it is easily generalizable to other applications of the Semantic essary to explore this knowledge base for inferring relations among Web. It is only necessary a correct conceptualization of the new application domain by an ontology containing specific instances, pose, the LIKO query language is introduced in our proposal. Some over which to apply the semantic reasoning process sketched here of its query operators are shown in Table 1 Our future work is focused on detecting new semantic relations 2. 4 Inference of semantic Relations and improving their ranking process, so that high quality recom mendations can be offered to the end users Our approach focuses on identifying several types of semantic relations among the instances contained in our knowledge base 4. REFERENCES path which consists of a set of instances joined by Object proper- [l] V. Geroimenko and C Chen. Visualizing the Semantic Web ties. For example, assume a Tv program presented by a person(in- Springer Verlag, 200 stance of the Credits class)who has also taken part in other content [2] Y Blanco, J.J. Pazos, A. Gi as director, as shown in the following chain(the properties appear Diaz, M. Lopez and B Barragans. AVATAR: An advanced Multi-Agent Recommender System of Personalized TV Contents by Semantic Reasoning. In Proceedings of 5th Quiz-Show hasPresenter Credits DirectorofComedy-Mowie International Conference on Web Information Systems Engineering, volume 3306 of Lectures Notes in Computer Even though there does not exist an explicit Object property Science, pages 415-421. Springer-Verlag, 2004 tween the quiz show he aforementioned comedy movie, it is 1079
Table 1: Query operators of the LIKO language Operators Input Output . A class defined in the ontology. Direct subclasses related to the input class. / A class defined in the ontology. Direct superclasses referred to the input class. >> A property of the ontology. Direct subproperties related to the input property. ⇐ Instances of classes defined in the ontology. Properties where the input class appears as domain. >⇒ Instances of classes defined in the ontology. Properties where the input class appears as range. To define specific information on which apply the precepts of our semantic reasoning approach, instances of the ontology classes have been included in this knowledge base. These instances are referred to specific TV programs extracted from databases that store abundant information about TV scheduling, such as www.imdb.com. Regarding the properties, note they establish explicit relations among the instances contained in the ontology. The reasoning process proposed in AVATAR takes as starting point these explicit relations and infers other implicit ones from them. 2.2 The User Profiles As we said in Sect. 2.1, our system must reason on the viewer preferences. For that reason, a mechanism for representing formally the user profiles is required. Our approach proposes to reuse the knowledge contained in the TV ontology for that purpose. So, we define the user profiles as OWL ontologies (which we will refer to as ontology-profiles), built incrementally as the system receives information about the user viewing behavior. This way, the ontology-profiles contain hierarchies of classes, specific instances and properties that are extracted from the TV ontology. In addition, the system also storesthe level of interest for a user, referred to each instance defined in his/her ontology-profile. A user modeling based on ontologies allows AVATAR to compare different profiles by applying the knowledge about the TV domain, beyond a simple syntactic comparison tackled in previous approaches. So, our recommender is able to find programs semantically similar to those ones that have been of interest for the user. 2.3 LIKO: A query language AVATAR takes as starting point the user ontology-profile to offer him/her a personalized recommendation. Once the instances de- fined in this profile have been located in the TV ontology, it is necessary to explore this knowledge base for inferring relations among these programs and the remaining ontology instances. For that purpose, the LIKO query language is introduced in our proposal. Some of its query operators are shown in Table 1. 2.4 Inference of Semantic Relations Our approach focuses on identifying several types of semantic relations among the instances contained in our knowledge base. We are working on a semantic relation named direct semantic path which consists of a set of instances joined by Object properties. For example, assume a TV program presented by a person (instance of the Credits class) who has also taken part in other content as director, as shown in the following chain (the properties appear in square brackets): Quiz-Show [ hasPresenter ] Credits [ DirectorOf ] Comedy-Movie Even though there does not exist an explicit Object property between the quiz show and the aforementioned comedy movie, it is not hard to see these contents are related, because the same person is involved in both of them. After the inference process, the goal is to rank the semantic paths detected by extracting the most relevant ones. The most signifi- cant relations involve the instances contained in the user ontologyprofile and other instances related to the former ones from a semantic point of view. We refer to this new relation as semantic nexus. So, there exists a semantic nexus between two instances of classes (or two properties) when these classes (properties) are siblings or are related by the subsumption principle (relations “is-a”). The relations defined here, which are easily discovered by the LIKO query operators shown in Table 1, support the proposed reasoning process and improve the personalized recommendations presented the users. 3. CONCLUSIONS AND FURTHER WORK In this paper a TV recommender system has been presented. The main contribution of our tool is the application of a semantic reasoning process on the descriptions of the programs and the user preferences. This approach allows to infer hidden relations in the system knowledge base, taking as starting point the instances contained in an OWL ontology about the TV domain, and the properties that define explicit relations among them. The existing approaches do not reason about the semantics of the offered programs. This way, the inference and reasoning processes are greatly hampered. Our proposal provides the users with varied and personalized recommendations, by discovering semantic relations among the programs of interest for them and other ones contained in our knowledge base. On the other hand, note that our proposal of semantic reasoning is not only valid for a TV recommender system like AVATAR, but also it is easily generalizable to other applications of the Semantic Web. It is only necessary a correct conceptualization of the new application domain by an ontology containing specific instances, over which to apply the semantic reasoning process sketched here. Our future work is focused on detecting new semantic relations and improving their ranking process, so that high quality recommendations can be offered to the end users. 4. REFERENCES [1] V. Geroimenko and C. Chen. Visualizing the Semantic Web. Springer Verlag, 2003. [2] Y. Blanco, J. J. Pazos, A. Gil, M. Ramos, A. Fernández, R. P. Díaz, M. López and B. Barragáns. AVATAR: An advanced Multi-Agent Recommender System of Personalized TV Contents by Semantic Reasoning. In Proceedings of 5th International Conference on Web Information Systems Engineering, volume 3306 of Lectures Notes in Computer Science, pages 415–421. Springer-Verlag, 2004. 1079