A Probabilistic Semantic Based Mixture Collaborative Filtering

Personalized recommendation techniques play more and more important roles for the explosively increasing of information nowadays. As a most popular recommendation approach, collaborative filtering (CF) obtains great success in practice. To overcome the inherent problems of CF, such as sparsity and scalability, we proposed a semantic based mixture CF in this paper. Our approach decomposes the original vector into semantic component and residual component, and then combines them together to implement recommendation. The semantic component can be extracted by topic model analysis and the residual component can be approximated by top values selected from the original vector respectively. Compared to the traditional CF, the proposed mixture approach has introduced semantic information and reduced dimensions without serious information missing owe to the complement of residual error. Experimental evaluation demonstrates that our approach can indeed provide better recommendations in both accuracy and efficiency.