Comparing the Effectiveness of Feature Analysis and Visual Analysis Machine Learning Approaches to Classifying Music Genres

Music genre classification remains a fundamental challenge in music information retrieval, with applications spanning automated music recommendation and content organization on streaming platforms. This study compares two prevalent machine learning approaches: feature-based analysis using Random Forest (RF) classifiers and visual analysis using Convolutional Neural Networks (CNNs). Using the GTZAN dataset containing 1,000 audio samples across 10 genres, we evaluate both methodologies through comprehensive performance metrics and cross-validation. The feature-based approach employs manually extracted audio features including Mel-Frequency Cepstral Coefficients (MFCCs), spectral centroid, chroma, and temporal characteristics. The visual approach processes mel-spectrograms through a CNN architecture optimized for small datasets via Global Average Pooling, reducing parameters from 5.3 million to 392,000. Results demonstrate that the optimized CNN achieves superior performance with 67.67% mean accuracy and 5.46% standard deviation in 5-fold cross-validation, compared to the RF model's 54.50% accuracy with 20% variance. While both approaches struggle with genres like disco and rock, the CNN approach shows more consistent classification across all genres. These findings suggest that visual analysis through properly configured CNNs outperforms feature-based methods for music genre classification, particularly when architectural adjustments account for limited training data.