Gated Probabilistic Matrix Factorization: Learning Users’ Attention from Missing Values / 1888
Shohei Ohsawa, Yachiko Obara, Takayuki Osogami
Recommender systems rely on techniques of predicting the ratings that users would give to yet unconsumed items. Probabilistic matrix factorization (PMF) is a standard technique for such prediction and makes a prediction on the basis of an underlying probabilistic generative model of the behavior of users. We investigate a new model of users' consumption and rating, where a user tends to consume an item that emphasizes those features that the user seeks to enjoy, and the ratings of the users are more strongly affected by those features than others. We incorporate this new user model into PMF and show that the resulting method, Gated PMF (GPMF), improves the predictive accuracy by several percent on standard datasets. GPMF is widely applicable, as it is trained only with the ratings given by users and does not rely on any auxiliary data.