Auto-Tagger of Contextual Activity Tags for Music Tracks

Resumen

In the music domain, tags are commonly used to categorize the different songs a user listens to on streaming services such as Spotify or Deezer. Different types of tags have been used, such as those related to the musical genre, others related to the emotions evoked by the song, and others related to the user's context of action or activity. These tags are usually provided by users in social networks and are often present in popular songs in the catalog. However, new songs added to the catalog or those belonging to the long tail do not have these tags and the need arises to create auto-taggers capable of tagging these songs. Several works and datasets have been proposed in the literature to solve this problem. In this paper we propose to infer activity context tags from simple song features obtained from the Spotify streaming service. A comparison of different multi-label classification models is presented and discussed. The results show how Spotify API features can be employed as simple features to use together with multi-label classification models to infer context labels for tracks. The results also point to the need for better representation of songs to use with these models to differentiate between closely related contexts.