Peningkatan Kualitas K-Means Clustering Data Audio Musik Menggunakan Transformasi TableDC

Abstract

Clustering audio music data with high features typically suffers from performance degradation due to the curse of high dimensionality. A dataset with 518 classical K-Means features typically struggles to model nonlinear relationships between data. The purpose of this study is to analyze the implementation of the TableDC latent space transformation technique in the preprocessing stage before K-Means on the FMA Small dataset. This case study contains 8,000 songs with 518 audio features and is divided into eight music genres. The performance of K-Means on the original data is compared with that of K-Means on the latent space extracted by TableDC. The analysis is performed using several metrics such as the Silhouette Score, Davies-Bouldin Index, Calinski-Harabasz Index, Adjusted Rand Index, inertia or WCSS, and the number of iterations. The experimental results indicate a percentage improvement offered by the method. The Silhouette Score increased by 53 percent from the initial value of 0.0249 to 0.0382. Similarly, the ARI value increased from the initial value of 0.0876 to 0.0893. However, these absolute values remain very low, indicating that the formed cluster structures are still weak and substantially overlapping. In this case, the latent representation contributed to increasing the convergence efficiency from 63 to 47 iterations. The WCSS value also decreased from 3,433,413 to 20,628. However, unlike the two previous indicators, the linear-based DBI and CHI actually obtained better results compared to the initial model, which demonstrates the model's weakness in the context of conventional evaluation. Overall, the TableDC transformation has been shown to improve computational efficiency, but its performance has not fully resolved the issue of overlapping class separation.