A3MR: Attentive Auto-encoder for Acoustic-assisted Music Recommendation

In the era of big data, users often seem to be unable to make choices in the face of such a huge range of options. This is especially the case with music; consequently, personalized music recommendation has attracted much interest. Many effective strategies have been offered by researchers, nevertheless, they still face two challenging problems: (1) how to model the complex usermusic relationships from sparse implicit feedback data, and (2) how to introduce auxiliary information to improve the recommendation effect. To copy with these challenges, we propose an Attentive Auto-encoder for Acoustic-assisted Music Recommendation (A(3)MR), which takes user historical behaviors, music acoustic features, and similar music to the objective music into account. Especially, we design a multi-attention layer to learn complex user-music relationships in order to gain the hidden representation of the objective music based on uses' behaviors. Besides, we use an embedding layer to generate music representation based on acoustic features, and cluster the similar music to objective music in order to predict users' preference. We conduct a series of experiments on the real-world dataset to evaluate the proposed model, and the results indicate that it is effective.