Feature selection vs. Feature Space Transformation in automatic music genre classification tasks

Automatic classification of music genres is an important task in music information retrieval research. Nearly all state-of-the-art music genre recognition systems start from the feature extraction block. The extracted acoustic features often could tend to be correlated or/and redundant, which can cause various difficulties in the classification stage. In this paper we present a comparative analysis on applying supervised Feature Selection and Feature Space Transformation algorithms to reduce the feature dimensionality. We discuss pros and cons of the methods and weigh the benefits of each one against the others.