With advent of smart TV and social TV services, TV users can consume not only TV programs but also various web contents. Therefore, for smart TV services, it is natural to recommend the related web contents to TV viewers while they are viewing TV programs. In order to do so, it is required that (i) personalized recommendation be possible; (ii) grouping of similar taste users be necessary; and (ii) recommendation of web contents that are highly semantically related with TV program contents be possible. In this dissertation, we propose a unified topic model for similar taste user grouping and cross-domain recommendation between TV and web domains based probabilistic association between TV programs and web contents. Our proposed unified topic model is based on three Latent Dirichlet Allocation (LDA) models: the first one is a topic model of TV users, the second one is a topic model of the description words for watched TV programs, and the last one is a topic model of description words for web video contents. Our unified topic model identifies the semantic relation between TV user groups and description word groups so that more meaningful TV program recommendations and cross-domain recommendation can be made. The unified topic model also overcomes an item ramp-up problem such that new TV programs and new web contents can reliably be recommended for TV users. In order to verify our proposed method of unified topic modeling based TV user grouping, TV program recommendation and cross-domain web content recommendation for social TV services, we use real TV watching history data, EPG data for seven months collected by a TV poll agency and web content description data from YouTube. The experimental results show that the proposed unified topic model yields an average 81.4% precision for 50 topics in TV program recommendation.
본 논문은 유사 시청 사용자 그룹핑과 TV 프로그램 콘텐츠 추천 및 TV 프로그램 콘텐츠와 웹 콘텐츠의 cross-domain 추천을 위한 통합 토픽 모델을 제안한다. 제안 기술은 토픽 모델링 기법 중 Latent Dirichlet Allocation(LDA) 방법을 기반으로 하여 서로 다른 세 종류의 데이터 (시청 TV 프로그램 데이터, TV 프로그램 description 데이터, 웹 콘텐츠 description 데이터)를 통합된 토픽 모델로 구성하여 하나의 프레임워크 내에서 은닉 토픽들을 생성하며 다양한 추천이 가능하도록 하였다. 추천 알고리즘의 성능평가를 위해 실제 TV 시청기록 데이터를 이용하여 추천 정확도를 측정하였으며 실제 테스트 사용자들을 모집하여 사용자 평가를 통해 웹 콘텐츠 cross-domain 추천 성능을 검증하였다.
