Top-N Recommendations by Learning User Preference Dynamics

Yongli Ren, Tianqing Zhu, Gang Li, Wanlei Zhou

Abstract: In a recommendation system, user preference patterns and the preference dynamic effect are observed in the user ×item rating matrix. However, their value has barely been exploited in previous research. In this paper, we formalize the preference pattern as a sparse matrix and propose a Preference Pattern Subspace to iteratively model the personal and the global preference patterns with an EM-like algorithm. Furthermore, we propose a PrepSVD-I algorithm by transforming the Top-N recommendation as a pairwise preference learning process. Experiment results show that the proposed PrepSVD-I algorithm significantly outperforms the state-of-the-art Top-N recommendation algorithms.