IPTV Viewing And Microblogging Behavior Models For TV Program Recommendation

Nowadays, living standards have improved so much, the number of TV programsare increasing rapidly. Viewers have to give much time and strength to select the rightTV programs from the massive data. It is necessary for the service providers to rec-ommend the programs for viewers. The trend is towards the personalized televisionservice. The explosive growth of microblogging, social network and IPTV serviceshas opened up new opportunities for service providers to understand user activities andpreference.In this study, we analyze the TV program recommendation from two data sourcesincluding the discussion of TV on the microblogging platform and the IPTV records.Microblogging services present two advantages for recommendation:1. Friend-ship network helps to mimic the world-of-mouth recommendation in one’s real life;2. The rich user-generate-content reveals users’ interests as well as items’ properties.Based on probabilistic matrix factorization, we proposed a hybrid recommendationmodel with two regularizers: the social regularizer and the item similarity regular-izer. The social regularizer comes from the friendship and retweet behaviors betweenusers. Theitemsimilarityregularizercomesfromthetextsimilaritybetweenprograms.When validating our algorithm with the application of TV show recommendation onSina Weibo, the proposed algorithm is shown to outperform the state-of-the-art collab-orative fltering method by9%at most. In addition, we show that the hybrid model isrobust in the scenario of making recommendation for new users, a typical cold startsetting.One key advantage of IPTV over the traditional TV is that it makes the TV view- ing experience more interactive and personalized. The service providers can analyzethe records to build the TV programs recommender system. Compared to the Internetapplication, the recommendation for IPTV faces more challenges:1. A family has oneor several members;2. Each member’s taste can be generated using a mixture of in-terests;3. Each member tends to watch TV in certain time periods every week. Basedon above assumption, we propose a novel coupled LDA model, which considers topicof TV programs viewed as well as the timestamps of the viewing behaviors, in orderto capture the viewing patterns of the users along the topic as well as the time dimen-sions. We perform an in-depth study on several intrinsic characteristics of IPTV useractivities by analyzing the real data collected from an operational nation-wide IPTVsystem. Our analysis has revealed that the coupled LDA model is able to reveal tem-poral patterns in TV watching behaviors as well as coherent interest groups. And ourcoupled LDA model can perform a better recommender system.