Attentive Auto-encoder for Content-Aware Music Recommendation

Mobile Internet allows consumers the access to all musics on the mobile platform. Since it is not feasible to manually select music due to the size constraint of mobile devices, music recommendation has become a popular research topic in recent years, and researchers have proposed many effective methods such as collaborative filtering. However, with the tremendous increase of mobile users and music resources, customized music recommendations still face two challenges: (1) how to model complicated relations from user-music interaction data, and (2) how to integrate heterogeneous content information of musics. In this paper, we propose an Attentive Auto-encoder for Content-Aware Music Recommendation (\(A^2CAMR\)), which effectively integrates user behaviour records, music content, and similar musics of the target. In particular, we design a hierarchical attention-based encoder layer to learn fine-grained user-user and music-user relationships, thus produce behavior-based hidden representation of musics. We also employ an embedding layer to produce the content-based music representation, and cluster the similar music sets of the target music to predict users’ preferences in the decoder. We conduct extensive experiments on real-world dataset, and the results demonstrate the effectiveness of our model.