Clustering Music by Genres Using Supervised and Unsupervised Algorithms

This report describes classification methods that recognize the genres of music using both supervised and unsupervised learning techniques. The five genres, classical(C), EDM(E), hip-hop(H), jazz(J) and rock(R), were examined and classified. As a feature selection method, discrete Fourier transform (DFT) converted the raw wave signals of each song into the signal amplitude ordered by their frequencies. Based on the analysis of the characteristics of data set, final feature set were collected by averaging the amplitudes of corresponding two different frequency division (XL and XM ). For supervised learning, a training set (mtrain = 50/genre) was used to train the CART (Classification and Regression Tree), and the performance of the genre prediction by CART classifier was evaluated using a test set (mtest = 10/genre). A recognition rate of 86.7% for three genre classification (C, H, and R) was observed, and 60.7% for five genre classification (C, E, H, J, and R) was obtained. For unsupervised learning algorithm, K-means clustering was performed on an unlabeled set of data (m = 60/genre) to cluster the music into genres, and showed purity of 84.4% for three genre classification, and 62.0% for five genre classification.