Feature Selection for Music Emotion Recognition

Abstract

Feature selection is step in preprocessing that can be used to reduce data dimension and eliminate the irrelevan data. There are several algorithms in feature selection. This study will compare several feature selection algorithms, namely Sequential Forward Selection, Sequential Backward Selection, and Relief F to find features that are very influential in musical emotional recognition. The method in music emotion recognition uses Support Vector Machine with the RBF kernel. The experimental results show that based on the recognition results with the highest accuracy, the most influential features are the zero crossing rate, music mode, harmonics, pitch and energy obtained through the Sequential Backward Selection algorithm. The selection of features in this study can increase accuracy up to 8%.