Enhanced similarities for a music Recommender System

Abstract

Music is an essential part of human life. It is a way to express ourselves. Its importance gave rise to music Recommender Systems (RS) through e-commerce applications. The most commonly used technique in such applications is Collaborative Filtering (CF) that uses the User-Item (UI) matrix. The latter codes the users' preferences for items. The advantage of using CF methods is their simplicity and their relatively high efficiency. However, data sparsity deteriorates their efficiency. In this paper, we propose a process that can enhance the similarities among users in a way that can address the sparsity problem. The proposed user based CF algorithm alters the values of similarity matrix between users by incorporating a graph based method, improving the performance of the RS. We also propose a method that groups users using spectral clustering and together with the graph-based similarity method yield even more accurate predictions. Experiments have been conducted on a music dataset, highlighting the superiority of the proposed method against typical user-based CF.