An Empirical Study on the Effectiveness of Feature Selection and Ensemble Learning Techniques for Music Genre Classification

Abstract

Classical machine learning has long been utilized for classification and regression tasks, primarily focusing on tabular data or handcrafted features derived from various data modalities, such as music signals. Music Information Retrieval (MIR) is an emerging field that seeks to automate the management process of musical data. This paper explores the potential of employing ensemble learning techniques to enhance classification performance while assessing the impact of feature selection methods on accuracy and computational efficiency across three publicly available datasets: Spotify, TCC_CED, and GTZAN. The Spotify and TCC_CED datasets contain high-level musical features, such as energy, key, and duration, while the GTZAN dataset incorporates low-level acoustic features extracted from audio recordings. The empirical experiments and qualitative analysis reveal a significant performance improvement when employing ensemble learning techniques for handling high-level features. Furthermore, the findings suggest that applying appropriate feature selection methods can substantially reduce computational time. As a result, by strategically combining optimal feature selection and classification models, the performance can be boosted in terms of accuracy and computational time. This study provides insights for optimizing music genre classification tasks through the strategic selection and balancing of model performance, ensemble learning techniques, and feature selection methods, ultimately contributing to advancements of musical genre classification tasks in MIR.