Music Genre Classification Using Feature Subset Search

Jihae Yoon, Hyunki Lim, Dae‐Won KimInternational Journal of Machine Learning and Computing 2016 · DOI

Abstract

With the growing number of digital music, the automatic genre recognition problem has been receiving the spotlight in music retrieval information field. A large number of musical acoustic features are reported to degrade the genre classification performance and lead to heavy computational cost. In this paper, we propose a new method for selecting genre-discriminative feature subset from a large number of musical features. We show that the proposed method is able to improve the genre recognition accuracy compared to the traditional selection method.

Introduction

In the field of Music Information Retrieval (MIR), researchers can extract nearly a thousand acoustic features from a single audio track—ranging from rhythmic patterns to harmonic structures. However, this abundance of data presents two major hurdles: the "Curse of Dimensionality," where irrelevant or redundant features degrade model performance, and high computational costs. Processing hundreds of features requires significant time and memory, and existing selection methods like mRMR often suffer from high time complexity ($O(N^2)$). This paper introduces an efficient feature selection method that balances classification accuracy with computational speed by extending the "min-redundancy max-relevance" (mRMR) algorithm using conditional mutual information (CMI).

Methodology

The proposed approach utilizes an incremental search method based on Conditional Mutual Information (CMI) to refine the feature selection process. Unlike standard mRMR, this method aims for a complexity closer to $O(N)$ through a multi-step process:

  1. Initial Sorting: Features are ranked by their individual dependency on the target genre.
  2. CMI Selection: The algorithm iterates through the sorted list, adding features only if they provide significant new information not already captured by the existing subset.
  3. Hybrid Fallback: To ensure a complete subset, the algorithm falls back to standard mRMR if the CMI search does not yield a sufficient number of features.

Experimental Setup

The study evaluated the method using a custom dataset of 115 songs across five genres: Rap, Electronic, Jazz, Metal, and Classic. Features were extracted from the first 30 seconds of each song using two industry-standard toolboxes:

  • MIRtoolbox: 369 features extracted.
  • JAudio: 916 features extracted.

Results and Discussion

The experimental results demonstrate that "less is more" in music genre classification. Using a compact subset of 50 features often outperformed the full set of over 900 features, confirming that feature noise is a primary inhibitor of accuracy.

Accuracy and Stability

The proposed method showed reliable improvements across different classifiers, including Naïve Bayes (NB) and k-Nearest Neighbor (k-NN). For instance, using the JAudio feature set and a Naïve Bayes classifier, the method improved accuracy from a 45% baseline (using all features) to approximately 64%.

Accuracy comparison for Naïve Bayes classifier showing the baseline vs. proposed method.

Furthermore, the method maintained high performance across different feature sets and classifiers. In tests using the JAudio dataset with a Neural Network (NN), the proposed method consistently outperformed traditional selection tools like ReliefF, which often fell below the baseline.

Accuracy for JAudio features using a Neural Network classifier.

Computational Efficiency

Beyond accuracy, the proposed method significantly reduced processing time. Selecting 30 features from the JAudio set took 28.31 seconds, compared to 40.99 seconds for the standard mRMR algorithm, representing a substantial reduction in computational overhead.

Conclusion

This research highlights the effectiveness of CMI-based feature selection in improving both the accuracy and efficiency of music genre classification. While the current study utilized a relatively small dataset, the results suggest that reducing feature redundancy is key to robust MIR systems. Future work will focus on scaling the approach to larger datasets and refining the hybrid fallback mechanism to maintain $O(N)$ complexity throughout.

Summary updated

Abstract

With the growing number of digital music, the automatic genre recognition problem has been receiving the spotlight in music retrieval information field. A large number of musical acoustic features are reported to degrade the genre classification performance and lead to heavy computational cost. In this paper, we propose a new method for selecting genre-discriminative feature subset from a large number of musical features. We show that the proposed method is able to improve the genre recognition accuracy compared to the traditional selection method.
BibTeX
@article{yoon2016music,
  author = {Jihae Yoon and Hyunki Lim and Dae‐Won Kim},
  title = {{Music Genre Classification Using Feature Subset Search}},
  journal = {International Journal of Machine Learning and Computing},
  year = {2016},
  doi = {10.18178/ijmlc.2016.6.2.587},
}
Figure