Klasifikasi Genre Musik Menggunakan Metode Hibridasi Particle Swarm Optimization dan Backpropagation

Abstract

Automatic music genre classification is one of the fundamental challenges in the field of Music Information Retrieval (MIR) due to the high similarity of acoustic characteristics among music genres. This study proposes a hybrid approach that combines Mel-Frequency Cepstral Coefficients (MFCC) for acoustic feature extraction with a Backpropagation Neural Network (BPNN), whose hyperparameters are optimized using Particle Swarm Optimization (PSO). The GTZAN dataset, consisting of 1,000 audio files across 10 music genres, was used in this study. The PSO optimization process automatically searches for the optimal values of five BPNN hyperparameters, namely the number of neurons in three hidden layers, the learning rate, and the dropout rate, using a 5-fold cross-validation scheme. The evaluation of PSO control parameters (the cognitive coefficient c1, social coefficient c2, and inertia weight w) indicates that the combination of c1 = 1.2, c2 = 1.5, and w = 0.7 achieves the lowest average validation loss of 0.9463 and is therefore selected as the optimal PSO configuration. Based on this configuration, PSO identifies the optimal BPNN hyperparameters as 192 neurons in the first hidden layer, 192 neurons in the second hidden layer, 32 neurons in the third hidden layer, a learning rate of 0.000185, and a dropout rate of 0.379. The resulting PSO-BPNN model achieves an average accuracy of 84.40%, precision of 85.09%, recall of 84.40%, and F1-score of 84.34%, outperforming the baseline BPNN model, which achieves an accuracy of 82.80%, precision of 83.60%, recall of 82.80%, and F1-score of 82.80%. Among the evaluated genres, classical achieved the highest classification performance, followed by metal, blues, and jazz, whereas rock, disco, and reggae consistently exhibited the lowest classification performance due to their acoustic similarities with other genres. These findings demonstrate that PSO-based optimization effectively improves the performance of BPNN for MFCC-based music genre classification, although the overall improvement remains relatively modest in absolute terms