Similarity Based Matrix Factorization for Recommender Systems

Gen Zhang,Xu Zhou
DOI: https://doi.org/10.1109/iscid.2017.22
2017-01-01
Abstract:In the area of computational intelligence, recommender systems play important roles in both commercial and social activities. Among its implement methods, matrix factorization is one of the most popular approach with various merits. However, traditional matrix factorization methods either roughly aggregate it into final task or ignore users/items similarity, which may easily introduce the cause of over-fitting or under-fitting. To tackle this problem, in this paper, we propose a similarity based matrix factorization for recommender systems by considering both global and local information within and between users and items. Specifically, it applies a novel method to extract the latent information within users and items, and captures the local similarity by a low-dimension manifold. Then, global and local information are integrated through a manifold regularized matrix factorization. Finally, the reconstructed method can be used in recommender systems. Experiments on two benchmark data sets illustrate the outstanding performance of our method in terms of prediction accuracy.
What problem does this paper attempt to address?