Exploiting and Exploring Hierarchical Structure in Music Recommendation

Abstract

Collaborative Filtering (CF) approaches have been widely applied in music recommendation as they provide users with personalized song lists. However, CF methods usually suffer from severe sparsity problem which greatly affects their performance. Previous works mainly use music content information and other external resources to relieve it, while they ignore that music entities are multi-typed and items are tied together within a hierarchy. E.g., for a track, we can identify its album, artist and associated genres. Therefore, in this paper, we propose a framework which utilizes the hierarchical structure in two ways. On one side, we exploit the hierarchical links to find more reliable neighbors; On the other side, we explore the effect of hierarchical structure on users’ potential preferences. In a further step, we incorporate the two aspects seamlessly into an integrated model which could make use of the advantages of both sides. Experiments conducted on the large-scale Yahoo! Music datasets show: (1) our approach significantly improves the recommendation performance; (2) compared with baselines, our approach is much more powerful on the even sparser training data, demonstrating that our approach could effectively mitigate the sparsity issue.