Current music recommendation models can mine the general preference between users and music

which are unable to distinguish the differential preference of different users towards the same song. Therefore

we propose a hierarchical attention representation model (HARM) to improve the music recommendation quality. HARM utilizes attributes of users and contents of music to learn music representation from the perspective of multi-dimension and mine the preference relationship between user and music. In order to mine users' differential preference on music's multi-field feature

a user feature-dependent attention network is designed. In addition

in order to mine the impacts of different historical behavior on user preference and learn sequential dependency of user's

a music-dependent attention network is designed. Finally

recommendation is generated by using a softmax function to calculate the preference distribution of users on candidate songs. The experimental results on 30Music and MIGU datasets shows that

comparing with the existent recommendation models

HARM can gain significant improvement on Recall and MRR.