Automatic playlist generation using convolutional neural networks and recurrent neural networks

Nowadays, a great part of music consumption on music streaming services are based on playlists. Playlists are still mainly manually generated by expert curators, or users, process that in several cases is not a feasible with huge amount of music to deal with. There is the need of effective automatic playlist generation techniques. Traditional approaches to the problem are based on building a sequence of music pieces that satisfies some manually defined criteria. However, being the playlist generation a highly subjective procedure, to define an a-priori criterion can be an hard task in several cases. In this study we propose an automatic playlist generation approach which analyses hand-crafted playlists, understands their structure and evolution and generates new playlists accordingly. We adopt Recurrent Neural Network (RNN) for the sequence modelling. Moreover, since the representation model adopted to describe each song is determinant and is also connected to the human perception, we take advantages of Convolutions Neural Network (CNN) to learn meaningful audio descriptors.