Collaborative filtering based on rating psychology

Nowadays, products are increasingly abundant and diverse, which makes user more fastidious. In fact, user has demands on a product in many aspects. A user is satisfied with a product usually because he or she likes all aspects of the product. Even only few of his or her demands or interests did not be satisfied, the user will have a bad opinion on the product. Usually, user's rating value for an item can be divided into two parts. One is influenced by his or her rating bias and other user's rating for the item. The other is determined by his or her real opinion on the item. The process of rating an item can be considered as an expression of user's psychological behavior. Based on this rating psychology, a novel collaborative filtering algorithm is proposed. In this algorithm, if one latent demand of the user is not satisfied by the item, the corresponding rating value will be multiplied by a penalty value which is less than 1. The parameters in the model are estimated using stochastic gradient descent method. Experiment results show that this algorithm has better performance than state-of-the-art algorithms.