TV program data has been increased significantly in recent years, which makes it difficult TV viewing environment. It is important for TV users to easily obtain their preferred contents (items) from excessive amounts of TV program contents available at TV terminal sides. Matrix Factorization (MF) is known as a powerful technique, which is one of the collaborative filtering methods for item recom-mendation. Basically, the MF based methods has popularly been used in movie recommendation sys-tems where user’s rating data on movie contents are usually available. However this is not a common case in TV recommendation applications because TV recommendation systems can only access the implicit data with browsing history, watching history and so on. The accurate computation of rating values from user’s watching history of TV programs is essential for TV program recommendation. Therefore, the basic MF cannot be directly applied for TV program recommendation and cannot cap-ture the temporal changes in user preference for watching TV programs. The basic MF learns all the data at one time. In this thesis, we propose a rating computation model and a recommendation system with a hybrid MF update method based on a probabilistic MF framework. The proposed method of estimating the rating value reflect user watching history with the frequency of watching TV program series. Also, the hybrid MF update method can capture the dynamics of the time-varying user prefer-ence based on the off-line and on-line update. For test, we propose a top-k list model of MF based on the user preference on TV program genres and the watching history. Our model outperforms the basic MF model in precision performance with 80.4% precision accuracy.