Feature selection and evaluation for genre classification of symbolically encoded classical music with the aid of machine learning

Abstract

This work defines useful features for the classification of symbolically encoded music into 14 classical genres namely chorale, symphony, etude, fugue, prelude, contrafactum, sonata, mazurka, motet, sonatina, waltze, concerto, Gregorian chant and scherzo. Features are based on Music Theory and grouped into seven categories: distances in the harmonic mobius strip, distances on the line of fifths, scale, rhythmic syncopation and meter, polyphony measurements, duration and instrumentation. Features are extracted and ranked combining 5 filter-based methods. Six Machine Learning algorithms are defined for classification: three Support Vector Machines, one Bayesian network, the C4.5 and random forests. Using nested cross-validation for training and testing and considering all the features, the Bayesian network classifier yields 84.10% empirical accuracy. The FEATUROMETRE process measures the usefulness of the feature subsets in an approach similar to wrapper methods, conveying relevant information to domain experts. Another experiment measures the usefulness and accuracy of features individually and by category using FEATUROMETRE. Grouping the music pieces by their period, the measured accuracy with the random forest classifier in the second experiment reaches 89.81%.