Feature selection and feature learning in arousal dimension of music emotion by using shrinkage methods

Music emotion recognition is an important topic in music information retrieval area. A lot of acoustic features are used to train a music classification or regression emotion model. However, these existing features may not be efficient for classification or regression task. Furthermore, most works do not explain why these features do work for classification. In our work, eight features are extracted to represent the arousal dimension of music emotion, and various commonly used statistical learning methods such as Logistic Regression, and tree-based methods are applied to interpret important features. Then the shrinkage methods are applied to feature selection and classification in music emotion recognition for the first time. Our tests show that the proposed approaches are efficient for feature selection just as entropy-based filter methods, and better than wrapper methods. The shrinkage methods can produce more continuous and low variance model than wrapper methods. Then, we discover that the most useful features are low specific loudness sensation coefficients (low-SONE), root mean square and loudness-flux. Moreover, the shrinkage methods apply in logistic regression perform better for classification than most of other methods. We get an average accuracy rate of 83.8 %.