Analyzing and enhancing music mood classification : an empirical study

In the computer age, managing large data repositories is one of the common challenges, especially for music data. Categorizing, manipulating, and refining music tracks are among the most complex tasks in Music Information Retrieval (MIR). Classification is one of the core functions in MIR, which classifies music data from different perspectives, from genre to instrument to mood. The primary focus of this study is on music mood classification. Mood is a subjective phenomenon in MIR, which involves different considerations, such as psychology, musicology, culture, and social behavior. One of the most significant prerequisitions in music mood classification is answering these questions: what combination of acoustic features helps us to improve the accuracy of classification in this area? What type of classifiers is appropriate in music mood classification? How can we increase the accuracy of music mood classification using several classifiers? To find the answers to these questions, we empirically explored different acoustic features and classification schemes on the mood classification in music data. Also, we found the two approaches to use several classifiers simultaneously to classify music tracks using mood labels automatically. These methods contain two voting procedures; namely, Plurality Voting and Borda Count. These approaches are categorized into ensemble techniques, which combine a group of classifiers to reach better accuracy. The proposed ensemble methods are implemented and verified through empirical experiments. The results of the experiments have shown that these proposed approaches could improve the accuracy of music mood classification