Improving GMM classifiers by preliminary one-class SVM outlier detection: Application to automatic music mood estimation

Abstract

Automatic estimation of music mood has emerged as an important task in Music Information Retrieval. It has direct applications in music search engines and cross-modal multimedia tools. During the last years, Gaussian Mixture Models (GMM) became one of the most popular classifiers for mood estimation. One of the remaining key challenges is the impossibility to collect representative training data sets. With GMM classifiers, "unknown" test data can result in low log-likelihoods for all mood classes, so that the resulting decision becomes immethodical. Thus, we suggest using a preliminary outlier detection based on one-class Support Vector Machines (SVM). In this paper we introduce a novel approach to optimize the one-class SVM parameters via minimizing the differences between the fraction of outliers, fraction of support vectors and parameter v. © 2010 Springer-Verlag Berlin Heidelberg.