Personalized and context-aware TV program recommendations based on implicit feedback

This thesis presents a context aware recommender system based on implicit-only feedback for a televisive run-time environment. The dataset used in this thesis comprehends the past behavior of 7921 users on 119 TV channels broadcasted both over-the-air (digital terrestrial television - DTT) and by satellite (SAT) including free and pay-tv channels. First of all we studied users’ history and modelled their behavior adding contextual information as suggested in the literature on context- aware recommender systems. Then we developed a possible system architecture and tested three different algorithms to perform run-time context-aware shows rec- ommendation to the users. The first algorithm is a trivial one, non-contextual, based only on considerations about timing. The second algorithm is based on a contextual pre-filtering and then a recommendation based on timing. The last and third model is the real contextual one and is based on the users’ models we built. We tested the algorithms against the complete test set and against two tailorings of it, the first one on the most active users and the second one eliminating the main- stream channels. We discovered which are the contextual dimensions that improve the recommendation among familiar context, day and time-band and finally we studied how mixing them with channels and genres from users’ history affects it.