A Cross Domain Collaborative Filtering Algorithm Based on Latent Factor Alignment and Two-Stage Matrix Adjustment

Abstract

Sparsity is a tough problem in a single domain Collaborative Filtering (CF) recommender system. In this paper, we propose a cross domain collaborative filtering algorithm based on Latent Factor Alignment and Two-Stage Matrix Adjustment (LFATSMA) to alleviate this difficulty. Unlike previous Cross Domain Collaborative Filtering (CDCF) algorithms, we first align the latent factors across different domains by pattern matching technology. Then we smooth the user and item latent vectors in the target domain by transferring the preferences of similar users and the contents of similar items from the auxiliary domain, which can effectively weaken the effect of noise. Finally, we convert the traditional UV decomposition model to a constrained UV decomposition model, which can effectively keep the balance between under-fitting and over-fitting. We conduct extensive experiments to show that the proposed LFATSMA algorithm performs better than many state-of-the-art CF methods.