Music genre classification using similarity functions

We consider music classification problems. A typical machine learning approach is to use support vector machines with some kernels. This approach, however, does not seem to be successful enough for classifying music data in our experiments. In this paper, we follow an alternative approach. We employ a (dis)similarity-based learning framework proposed byWang et al. This (dis)similarity-based approach has a theoretical guarantee that one can obtain accurate classifiers using (dis)similarity measures under a natural assumption. We demonstrate the effectiveness of our approach in computational experiments using Japanese MIDI data.