Fast content independent playlist generation for streaming media

Playlist generation is an important part of online streaming media platforms. Different methods rely on different information to generate playlists. Approaches trained exclusively on previous playlists (interaction data) are content independent, thus they can work with playlists containing different content types. The data types, playlist generation methods and evaluation approaches are described in the first part of the paper. A novel method with similar performance to k Nearest Neighbour is proposed in second part. The method is based on random walk and returns tracks with the same probability distribution as a k Nearest Neighbour method would do with a special similarity metric and including all neighbours. As there is no need for neighbour search the runtime complexity is lower by orders of magnitude. The performance is comparable to playlist generation methods also incorporating artist meta-data (CAGH, SAGH).