An Empirical Study on Structured Dichotomies in Music Genre Classification

Ensemble learning approaches have been gaining popularity in the non-trivial task of multi-class classification. Some of these, including 1-against-1, 1-against-all, and dichotomy-based methods, are based on decomposing the class space of a multi-class task into a set of binary-class ones. In this work, we investigate whether they could help improve genre classification in music. In particular, we explore various dichotomy structures of binary classifiers in music data. In addition to the existing ones, we propose several strategies to build new binary tree structures. We base our approach on the observation that people find it easy to distinguish between certain classes and difficult between others. In our investigation, we use several base classifiers that are common in the literature and conduct series of empirical experiments on two benchmarking music datasets. We report the initial results of our investigation in this paper.