An Association-based Approach to Genre Classification in Music

Music Information Retrieval (MIR) is a multi-disciplinary research area that aims to automate the access to largevolume music data, including browsing, retrieval, storage, etc. The work that we present in this paper tackles a nontrivial problem in the field, namely music genre classification, which is one of the core tasks in MIR. In our proposed approach, we make use of association analysis to study and predict music genres based on the acoustic features extracted directly from music. In essence, we build an associative classifier, which finds inherent associations between content-based features and individual genres and then uses them to predict the genre(s) of a new music piece. We demonstrate the feasibility of our approach through a series of experiments using two publicly available music datasets. One of them is the largest available in MIR and contains real world data, while the other has been widely used and provides a good benchmarking basis. We show the effectiveness of our approach and discuss various related issues. In addition, due to its associative nature, our classifier can assign multiple genres to a single music piece; hopefully this would offer insights into the prevalent multilabel situation in genre classification.