Bipartite graph link prediction method with homogeneous nodes similarity for music recommendation

Abstract

Through analyzing users’ listening records, personalized music recommendation can not only help users find interesting music, but also help related enterprises improve user loyalty. This paper proposes an improved music recommendation method based on bipartite graph link prediction with homogeneous nodes similarity. Firstly, users’ music preference relations are expressed as positive and negative binary preference relations by the Complex Representation-based Link Prediction (CORLP) method, which improves the limitation of traditional recommendation method that can only represent unary preference relations. Secondly, the new method improves the CORLP method by attribute extraction and similarity calculation of homogeneous nodes including user nodes and music nodes. Thirdly, a new dataset based on the practical data from Shrimp Music Community is collected for facilitating the music recommendation task. The first-class music genres and second-class music genres of users are extracted by web crawling technology to calculate the similarity between user nodes. The rhythm and tempo are extracted by open source software to calculate the similarity between music nodes. Finally, the Top-N experiment is used to prove the performance of the proposed method compared with CORLP. In addition, the results reveal several new findings. Firstly, performance is significantly improved when the homogeneous nodes similarity is taken into account. Secondly, recommendation method with user nodes similarity shows a better performance compared with recommendation method with music nodes similarity.