Clustering of Musical Genres

The aim of this paper is to better understand the landscape of musical genres. We use public tag data (folksonomies) to these discover genres, and make use of APIs from the music site Last.fm gather tag data for songs and artists. To be comprehensive, a uniform sampling of tag documents will be used. API calls are slow, however, while there are about 100 million artists on the database; so we sample with a Markov chain that converges to a unifrom stationary distribution. With a corpus of tag multisets, each associated with an artist, there are several ways to define a genre. I explore simple co-occurrences, lower-rank representations of tags (Latent Semantic Indexing), and probability distributions of tags (Latent Dirichlet Allocation). For each of these we can define a distance function, respectively: relative frequency, cosine similarity, and Hellinger distance. Once distances have been defined we can cluster the genres into either flat or nested clustering. For flat clustering, I show how Kohenen’s Self Organizing Maps and correlation clustering via genetic algorithms can be used to show spatial relations between genres where K-means fails. To generate a nested clustering I employ hierarchical clustering, with complete linkage.