Predicting genre labels for artists using FreeDB

This paper explores the value of FreeDB as a source of genre and music similarity information. FreeDB is a public, dynamic, uncurated database for identifying and labelling CDs with album, song, artist and genre information. One quality of FreeDB is that there is high variance in, e.g., the genre labels assigned to a particular disc. We investigate here the ability to use these genre labels to predict a more constrained set of “canonical” genres as decided by the curated but private database AllMusic (i.e. multi-class learning). This work is relevant for study in music similarity: we present an automatic, data-driven method for embedding artists in a continuous space that corresponds to genre similarity judgements over a large population of music fans. At the same time, we observe that FreeDB is a valuable resource to researchers developing music classification algorithms; it serves as a reference for what music is popular over a large population, and provides relevant targets for supervised learning algorithms.