Generating Music Playlists with Hierarchical Clustering and Q-Learning

Automatically generating playlists of music is an interesting area of research at present, with many online services now offering “radio channels” which attempt to play through sets of tracks a user is likely to enjoy. However, these tend to act as recommendation services, introducing a user to new music they might wish to listen to. Far less effort has gone into researching tools which learn an individual user’s tastes across their existing library of music and attempt to produce playlists fitting to their current mood. This paper describes a system that uses reinforcement learning over hierarchically-clustered sets of songs to learn a user’s listening preferences. Features extracted from the audio are also used as part of this process, allowing the software to create cohesive lists of tracks on demand or to simply play continuously from a given starting track. This new system is shown to perform well in a small user study, greatly reducing the relative number of songs that a user skips.