When and What Music Will You Listen To? Fine-Grained Time-Aware Music Recommendation

摘要

Music recommendation which helps displaying proper songs to proper users, has attracted growing attention in recent years. Current music recommendation systems are facing two key challenges: (1) users' listening habits with respect to time haven't been well studied; (2) there are far less ratings than listening records in music providing systems. In this paper, we strive to address the above two challenges. We investigate when and what music will a user listen to, and we propose a Fine-grained Time-aware Music Recommendation (FTAMR) model. To be specific, we improve recommendation qualities from two sides: the user and the song. From the user's side, we study users' listening time-behavior in a fine-grained way and explore their short-term listening habits (e.g. in a day) and long-term listening habits (e.g. in several months); From the song's side, considering the sparsity of ratings and the characteristics (e.g., popularity) of each song, we define the asymmetric co-recommendation probability from a song to another, and cluster songs according to co-recommendation probabilities instead of similarities. Given users' listening records with time stamps, we make recommendation based on their listening habits and songs' co-recommendation probabilities. To validate the effectiveness of the proposed FTAMR model, we study the Last.fm data set and conduct extensive experiments. The results show that our approach can provide better recommendations.