Granular methods in automatic music genre classification: a case study

Classification of music files by using the characteristics of the songs based on its genre is a very popular application of machine learning. The focus of this work is on automatic music genre classification based on granular computing methods (fuzzy rough, rough and near sets). We have proposed a modified form of supervised learning algorithm based on tolerance near sets (TCL 2.0) with a goal of exploring the scalability of the learning algorithm to a well researched music database composed of several genres. In the tolerance near set method, tolerance classes are directly induced from the dataset using the tolerance level epsilon and a distance function. We have compared the tolerance-based near set algorithm to a family of nearest neighbour (NN) algorithms based on fuzzy rough methods (FRNN) available in the WEKA platform. In terms of performance, the classification accuracy of TCL 2.0 is identical to the Bayesian Networks (BN) Algorithm, and comparable with the Sequential Minimal Optimization (SMO) Algorithm. However, the average classification accuracy of FRNN algorithms and the classical rough sets algorithm is better than TCL 2.0, BN and SMO algorithms. For this dataset, any accuracy over 90% is considered a very good classification accuracy which is achieved by all tested classifiers in this work.