Timbral and Semantic Features for Music Playlists. Accepted presentation at the Machine Learning for Music Discovery.

Music recommendation methods based on song- level features (either derived from audio con- tent or metadata), suffer from not being able to identify clear relations between music items and listeners, whose perception of the quality of a received recommendation is affected by a wider range of factors. This problem is par- ticularly severe for the task of generating mu- sic playlists. The analysis of song character- istics in hand-curated playlists exhibits large within-playlist variability, indicating that gener- ating music playlists or creating suitable continu- ations may be an ill-defined problem. In this pa- per we analyze two different features, based on either timbral or semantic descriptors of songs, for the task of predicting whether a song is suit- able or not for a playlist. Our empirical results on a dataset of hand-curated playlists indicate that features extracted from semantic descriptors are better suited for this task.