Embedding of Playlists for Music Recommendation by Optimizing Non-smooth Cost Function

In the present, there are millions of songs in the world. Some cloud-based music services (e.g. Youtube, Spotify and Apple Music) gather songs and provide a friendly user interface to consumers. As the appearance of the ser- vices, it can help consumers to explore the large set of songs. In order to satisfy the preference of consumers, a good recommendation method is nec- essary. There are many scholarly works focused on the related topic in music recommendation. To check the performance of different models are good or not, the evaluation metric is used in general. There are a lots of different met- rics in the information retrieval field. However, most of them are non-smooth. This cause a problem that there are a mismatch between the optimizing cost function and the evaluation metric. To overcome the problem, we introduce learning to rank to help to eliminate the mismatch. In this work, we embed each song with a d-dimension vector and use a ranking measure NDCG (Nor- mailized Discounted Cumulative Gain) as our evaluation metric. Given a song sequence, our model predict a ranking of next song right after the sequence. The experiments show that our model perform better than others. It indicates the ranking prediction of our model is more approximating the real ranking.