Attribute-Biased-SVD in Recommeder System

Recommender System helps users to find interesting items through analysis of users’ behavioral patterns. After nearly 20 years’ work on this area, researchers have put forward a variety of recommendation models and algorithms to fulfill this goal. Among them, SVD and other factorization techniques have been widely used since they were put forward. This paper first describes the definition and geometric meanings of SVD and the principles of Latent Factor Model (LFM). Then we integrate item attributes and the ratings of users on these attributes into LFM with weighted biases and give the detailed inference of the descent learning rules. In the end of this paper, its effectiveness is proved by experiments on Yahoo! Music data sets, which to some extent demonstrates the inclusiveness and pervasiveness of factorization techniques.