Exploring the Design Space of Symbolic Music Genre Classification Using Data Mining Techniques

This paper describes a method based on data mining techniques to classify MIDI music files into music genres. Our method relies on extracting high level symbolic features from MIDI files. We explore the effect of combining several data mining preprocessing stages to reduce data processing complexity and classification execution time. Additionally, we employ a variety of probabilistic classifiers and ensembles. We compare the results produced by our best classifier with those obtained by more complex state of the art classifiers. Our experimental results indicate that our system constructed with the best performing combination of data mining preprocessing components together with a Naive Bayes-based classifier is capable of outperforming other more complex ensembles of classifiers.

1 Introduction

Some music genre classification systems emulate the way humans proceed to perform this task. When asked to classify music we are commonly provided with a list of representative titles of the genre. One is then expected to gain an understanding of the genre by generalizing from the combination of properties that characterize these given titles. Classification of new music titles is performed by evaluating their similarity with respect to the other titles that we already know belong to a certain category. This is one feature of the inductive learning process and one example of the kind of problem that classification algorithms are designed to solve.