Skip to Main Content
Classification of musical audio signals according to expressed mood or emotion has evident applications to content-based music retrieval in large databases. Wrapper selection is a dimension reduction method that has been proposed for improving classification performance. However, the technique is prone to lead to overfitting of the training data, which decreases the generalizability of the obtained results. We claim that previous attempts to apply wrapper selection in the field of music information retrieval (MIR) have led to disputable conclusions about the used methods due to inadequate analysis frameworks, indicative of overfitting, and biased results. This paper presents a framework based on cross-indexing for obtaining realistic performance estimate of wrapper selection by taking into account the simplicity and generalizability of the classification models. The framework is applied on sets of film soundtrack excerpts that are consensually associated with particular basic emotions, comparing Naive Bayes, k-NN, and SVM classifiers using both forward selection (FS) and backward elimination (BE). K-NN with BE yields the most promising results - 56.5% accuracy with only four features. The most useful feature subset for k-NN contains mode majorness and key clarity, combined with dynamical, rhythmical, and structural features.