Local Probabilistic Matrix Factorization for Personal Recommendation
Wenping Ma, Yue Wu, Maoguo Gong, Can Qin, Shanfeng Wang · 2017
Matrix factorization is an efficient method to predict users' preference in recommender systems. Probabilistic matrix factorization (PMF) is an improved version of matrix factorization, and its assumption is more realistic. However PMF is prone to over-fitting. The algorithm local low-rank matrix approximation (LLORMA) performs well as it assumes the observed ratings matrix is local low-rank. Inspired by this, local probabilistic matrix factorization (LPMF) is proposed by introducing LLORMA to PMF. LPMF can find a certain number of local optimal points to estimate the parameters, and it will effectively alleviate over-fitting in every local model. By testing two benchmark datasets, MovieLens and Netflix, the experimental results demonstrate that the proposed algorithm could help PMF to find better parameters to a material extent.