Music Playlist Continuation by Learning from Hand-Curated Examples and\n Song Features: Alleviating the Cold-Start Problem for Rare and Out-of-Set\n Songs

Abstract

Automated music playlist generation is a specific form of music\nrecommendation. Generally stated, the user receives a set of song suggestions\ndefining a coherent listening session. We hypothesize that the best way to\nconvey such playlist coherence to new recommendations is by learning it from\nactual curated examples, in contrast to imposing ad hoc constraints.\nCollaborative filtering methods can be used to capture underlying patterns in\nhand-curated playlists. However, the scarcity of thoroughly curated playlists\nand the bias towards popular songs result in the vast majority of songs\noccurring in very few playlists and thus being poorly recommended. To overcome\nthis issue, we propose an alternative model based on a song-to-playlist\nclassifier, which learns the underlying structure from actual playlists while\nleveraging song features derived from audio, social tags and independent\nlistening logs. Experiments on two datasets of hand-curated playlists show\ncompetitive performance compared to collaborative filtering when sufficient\ntraining data is available and more robust performance when recommending rare\nand out-of-set songs. For example, both approaches achieve a recall@100 of\nroughly 35% for songs occurring in 5 or more training playists, whereas the\nproposed model achieves a recall@100 of roughly 15% for songs occurring in 4 or\nless training playlists, compared to the 3% achieved by collaborative\nfiltering.\n