Similarity-enhanced collaborative filtering

Abstract

Matrix factorization (MF) and neighborhood models are the two most successful approaches in the context of Collaborative Filtering (CF). Recently, graph-based models have emerged as an effective alternative despite little attention has been devoted to them. In this dissertation, we exploit collaborative data to improve MF and graph-based recommendation. We propose a graph-based approach that exploits three-step paths starting from the target user in the user-item bipartite graph and relies on the scoring functions, which exploit distributional aspects of the ratings given by users in order to extract their latent information. Likewise, we propose a MF approach that exploits both user-user and item-item similarities extracted from the raw user-item rating matrix. We experiment with several publicly available datasets against relevant baselines. The results attest the value of our methods as they provide significant gains in several settings.