(Xunta de Galicia) project PGIDIT05PXIC32204PN. Ricardo TubÃoâ , Rafael Soteloâ â , Yolanda Blancoâ , MartÃn Lópezâ , Alberto Gilâ , José Pazosâ and Manuel.
A TV-ANYTIME METADATA APPROACH TO TV PROGRAM RECOMMENDATION FOR GROUPS *
Ricardo Tubío†, Rafael Sotelo††, Yolanda Blanco†, Martín López†, Alberto Gil†, José Pazos† and Manuel Ramos† †
Dept. of Telematics Engineering, University of Vigo, Spain. †† University of Montevideo, Montevideo, Uruguay.
ABSTRACT The advent of Digital TV and Personal Digital Recorders promise to change the way people watch TV. The higher efficiency of digital coding will lead to increasing the number of contents offered to the user, demanding automatic tools for content recommendation. In the other hand, digital recorders will permit a non-linear consumption model, enabling the creation of (automatic) personalized schedules that combine the appealing contents for a concrete user. This paper presents a semantic approach to content recommendation for groups of people, starting from TV-Anytime descriptions of the contents to be broadcast. Index Terms— TV-Anytime, recommender systems, multiple audience 1. INTRODUCTION Digital television is nowadays being deployed all around the world, offering many advantages to end users such as improved quality of audio and video, interactivity, mobility and a higher efficiency that permits to increase the number of broadcast channels or enable high definition. In parallel, digital settop boxes with local storage are emerging (the so-called Personal Digital Recorder or PDR), being able to record hundreds of hours of video, (automatically) schedule recordings or even merge contents to compound a virtual channel. This will change the traditional linear nature of TV, multiplying the possibilities available to the final user who can be easily overwhelmed by this new landscape. In this context, it is necessary to develop agents that recommend programs to the users, thus improving their viewing experience. These agents employ several strategies to compare the broadcast contents to the users’ profiles and *
their usage history, providing them with recommendations or even a personal channel (using automated recording capability). Current proposals [1] are only focused on personal recommendations, based on the profile of a particular consumer. However, television viewing is very often a group experience because people watch TV with their families or friends. As a recommendation for an individual may not be adequate or optimal for all the members of the group, it is necessary to develop new recommendation engines for this task. Such engines must consider the personal profiles and usage history of all individuals of the group. These personal data may be available in the same PDR (in the case of families or home mates), or may be in different PDRs (each one belonging to each individual). So, a normalized way for the PDR to represent and exchange personal profiles is necessary to set this framework. Normalization of the PDR plays an important role to make possible its insertion in the market. It gives a framework for the different agents (broadcasters, carriers, content creators, manufacturers…) to develop their products and to interact. TV-Anytime (TV-A) is a specification that aims to make possible advanced audiovisual services based on the PDR. Particularly, TV-A normalizes rich metadata describing the content, user profiles and usage history. This metadata may be broadcast together with the content or be available by other ways, for example a third party web site. TV-A Phase 1 is mainly focused on unidirectional networks, while Phase 2 takes into account bidirectional aspects. For instance, Phase 1 allows the PDR to have multiple user profiles, while Phase 2 introduces the capability to exchange personal profiles. Particularly it allows: (i) providers to receive detailed and comprehensive data from a wide range of PDR devices from different users, (ii) PDRs to directly exchange profiles as needed, and (iii) consumers to "carry" their profiles and other personal data
Work funded by the Ministerio de Educación y Ciencia (Gobierno de España) project TSI2007-61599, by the Consellería de Educación e Ordenación Universitaria (Xunta de Galicia) incentives file 2007/000016-0, and by the Consellería de Innovación, Industria e Comercio (Xunta de Galicia) project PGIDIT05PXIC32204PN.
[2]. TV-A permits the development of web services that can retrieve user information, paving the way for PDR and service to communicate. Since personal information is sensitive, TV-A requires the user to specify who is authorized to request his personal profile, and provides the security mechanisms [3]. This permits the PDR to gather the users’ profiles needed for a group recommendation. In the rest of the paper we present a novel approach to content recommenders that addresses the problem of computing suggestions for groups of individuals. 2. RECENT WORK Some strategies on personal content recommenders for TV include expert systems, Bayesian techniques, content-based methods, collaborative filtering, decision trees, and others [1]. Recommender AVATAR [4] is a new approach focused on extending current syntactic search engines to include reasoning based on well-known techniques of the Semantic Web. Regarding TV-A, some recent work implements several audiovisual services based on this metadata. For example, an implementation of a TV-A compliant metadata processing engine for a PDR is described in [5] that provides personalized services around an Electronic Content Guide, including a personal channel, and with strong focus on an automatic extraction algorithm of user preferences. Authors of [6] introduce a method and content structure for personalized data broadcasting using the TV-A package model. A personalized TV system based on TV-A metadata model is proposed in [7] and implemented in an MHP settop box. The authors adopt the content-based filtering approach as the recommending mechanism, using fuzzy inference to automatically generate the users’ profiles from their usage history.
multiple dimensions in order to compute similarity. Regarding these dimensions, our current implementation works with four ontologies extracted from the classification schemes Intention, Format, Content and IntendedAudience defined in the TV-A standards. Then, we define a semantic similarity metrics to measure the adequacy of a candidate program for a given user by comparing that program to those stored in his/her profile. Given a candidate program P and a program Bi defined in the profile of a user U, our metrics quantifies their similarity by considering two components: the first one (named hierarchical similarity or SimHie (P,Bi)) measures their closeness in the content hierarchy. This closeness is computed by (1), where the depth of a program is its level in the content hierarchy, and LCA of two programs identifies their Lowest Common Ancestor in that hierarchy.
SimHie ( P, Bi ) =
depth ( LCA( P, Bi ))
(1)
Max ( depth ( P ), depth ( Bi ))
In fact, the hierarchical similarity between the user’s preferences and candidate programs will be computed for each one of the four hierarchies and later averaged. The second component (which is named inferential similarity and denoted by SimInf (P, Bi)) discovers associations between programs that share characteristics semantically related to each other, by using reasoning techniques employed in the Semantic Web. The shared characteristics can be completely equal or, in some cases, just belong to the same parent class. This approach leads to (2), where the DOI index (Degree Of Interest) is the interest of the user in the shared characteristics (SC). This level of interest (a real number in [0,1]) is automatically obtained from the specific ratings the user has assigned to the programs related to those characteristics.
3. GROUP RECOMMENDATIONS In order to elaborate personalized recommendations for a group of users watching TV together, we start from an OWL ontology that (i) describes TV programs, (ii) classifies them in a content hierarchy, and (iii) relates them to other programs through their semantic characteristics (attributes such as cast members, location, dates…). Actually, taking the most of the TV-Anytime capabilities, we have devised an enhanced ontological model that reflects the multidimensional content classification scheme of TV-Anytime into our hierarchy structure (there is an ontology for each every TV-Anytime classification scheme). The descriptions of the programs to be broadcast are received in TV-Anytime format including their classifications in these schemes (as created by content producers). The instances of the scheduled programs are added to the system database, linking and classifying them in one or several of our ontologies, enabling comparisons in
SimInf ( P, Bi ) =
1
# SC
∑ DOI ( SC ) # SC j
(2)
j =1
Once similarity between P and each program Bi in U’s profile has been computed, we obtain the components SimHie (P,U) and SimInf (P,U), finally combined by a configurable parameter α as shown in (3). SemSim ( P, U ) = α ⋅ SimInf ( P, U ) + (1 − α ) ⋅ SimHie ( P, U )
(3)
Bearing in mind the fact that television is often viewed by multiple audiences (e.g. families) rather than by individuals, we extend the previous single user metrics to computing similarity between programs and groups of viewers. This way, to compute the adequacy of a program P for a set of N users, we firstly measure the similarity
between P and each user Ui, and then all of them are combined according to (4). N
SemSim ( P, G ) = ∏ K i × SemSim( P , U i )
(4)
i =1
The constants Ki endow each user with a different importance if necessary. This equation will punish programs not appropriate for one of the users of the group. All user profiles are available for the recommender agent either because they already exist in the PDR or because of the capability to exchange personal profiles provided in TV-A Phase 2 [2]. 4. CONCLUSIONS AND FUTURE WORK
In this paper, a novel approach to compute TV content recommendations for a group of viewers has been presented. To this aim, the new capabilities of TV-Anytime Phase 2 regarding exchange of the user’s profiles between devices favors the development of interoperable consumer electronic equipment. In addition, multidimensional classification of contents based on TV-Anytime metadata schemes has been included which improves the accuracy of current similarity metrics. In the future, we plan to extend this framework to define group profiles and to consider more complex personalization scenarios in which, for instance, it is possible to prevent users from viewing programs belonging to categories forbidden for them (e.g. sex and violence when there are children). 5. REFERENCES [1] L. Ardissono, A. Kobsa, and M. Maybury, Personalized Digital Television, 1st ed., vol. 6. Kluwer Academic Publishing, 2004. [2] ETSI TS 102 822-6-3 V1.1.1 TV-Anytime. Part 6: “Delivery of metadata over a bi-directional network”; Sub-part 3: Phase 2 - Exchange of Personal Profile” [3] ETSI TS 102 822-7 V1.1.1 TV-Anytime; Part 7: Phase 1 - Bidirectional metadata delivery protection. [4] Y. Blanco, J. Pazos, M. López, A. Gil, M. Ramos, “AVATAR: an improved solution for personalized TV based on semantic inference”, IEEE Transactions on Consumer Electronics, vol. 52, pp. 223-231,2006. [5] L. HeeKyung, K. Jae-Gon, Y. Seung-Jun, J. Hong, “Personalized TV services based on TV-anytime for personal digital recorder”, IEEE Transactions on Consumer Electronics, vol. 51, pp. 885-892, 2005. [6] Y. Ho, H. Lee, J. Soo, J. Woo, “Study on Personalized Data Broadcasting Service using TV-Anytime Metadata”, Proceedings of 10th International Symposium on Consumer Electronics, pp. 1-6, 2006. [7] H. Zhang; S. Zheng; “Personalized TV program recommendation based on TV-anytime metadata”, Proceedings
of the 9th International Symposium on Consumer Electronics, pp. 242-246, 2005.