Modeling user and item biases with Gaussian distribution for collaborative filtering

The collaborative filtering approach to recommender systems focuses on learning predictive models of user preferences, interests and behavior from community data, that is, the behavior of other available users. Matrix Factorization (MF) based approaches have been proven to be efficient collaborative filtering algorithm for rating-based recommender systems. But existing MF algorithms have several disadvantages, including ignoring the distribution of the ubiquitous user and item biases. In this work we present an improved probabilistic matrix factorization (IPMF) algorithm and its graphical model. We analyzed the statistical pattern of user and item biases in the MovieLens dataset. The user and item biases are normally distributed. The improved model takes user and item preference biases into account, thereby building a more accurate model. Further accuracy improvements are achieved by extending this model with nonnegative user feature vectors. We evaluated these methods on the MovieLens dataset, and we show that our experimental results are better than those previously reported on this dataset.