Item Attribute-Aware Probabilistic Matrix Factorization for Item Recommendation

Abstract

Probabilistic matrix factorization based recommendation algorithm has been widely applied in industry and academia due to its effectiveness and efficiency in dealing with large-scale data sets. However, recommendation algorithms built upon probabilistic matrix factorization seriously suffer from cold start problem, e.g., they fail to accurately learn the latent features for new registered users or new added items, leading to poor recommendation quality. In this paper, we propose an improved probabilistic matrix factorization based recommendation algorithm by jointing item attribute information with probabilistic matrix factorization framework, named Item Attribute-aware Probabilistic Matrix Factorization (IAPMF). Item attribute are exploited to constrain the process of probabilistic matrix factorization, and derive the latent feature of the new item from its neighbors, which are similar to the target new item in terms of content. Experimental results show that our proposed recommendation algorithm is superior to the traditional methods.