Automatic Categorization of Electronic Music Genres

Abstract

In this research we have used three different machine learning approaches on the automatic categorization of electronic music genres. We used Spotify for the collection of the data set, together with their custom track features. The selection of algorithms consists of K-nearest neighbors (KNN), Support vector machine (SVM) and K-means. The comparison between the supervised methods (KNN & SVM) was done using confusion matrices. We achieved an accuracy 70.1% and 75.5% respectively. For the unsupervised method (K-means) the performance was measured using the purity (49.7%) and Silhouette score (0.283). Principal Component Analysis (PCA) was used to visualize the clustering of K-means. We compared this to the original visualization of the data set to find differences and similarities. After this com- parison, we came to the conclusion that unsupervised machine learning methods find different ways of genre categorization compared to our current classification. We can use these findings to improve our current ways of automatic genre classification. In addition to it, noticeable differences in both supervised and unsupervised categorization provide new grounds for more detailed comparison between certain genres.