Improving matrix approximation for recommendation via a clustering-based reconstructive method

Matrix approximation is a common model-based approach to collaborative filtering in recommender systems. Many relevant algorithms that fuse social contextual information have been proposed. They mainly focus on using the latent factor vectors of users and items to construct a low-rank approximation of the user-item rating matrix. However, due to data sparsity, it is difficult for current approaches to accurately learn every user/item vector, which may cause low-quality recommendations for some users or items. But we implicitly detect some similar users or items based on the distance among the vectors. In this paper, we take advantage of the implicit similarity to improve matrix approximation. Borrowing parts of ideas from CodeBook Transfer, we propose a reconstructive method that compresses low-rank approximation into a cluster-level rating-pattern referred to as a codebook, and then constructs an improved approximation by expending the codebook. Experiments on real life datasets demonstrate our method improves the prediction accuracy of the state-of-the-art matrix factorization and social recommendation models.