Performance Comparison of K-Means and Hybrid Hierarchical–Partitioning Methods for Clustering Efficiency

Abstract

Clustering is a fundamental technique in data analysis, particularly for exploring patterns in large-scale datasets. While K-Means is widely used for its simplicity and efficiency, its performance is highly sensitive to centroid initialization, which can affect both clustering quality and convergence speed. Hierarchical clustering methods, such as agglomerative and divisive approaches, provide more structured and deterministic initialization but incur higher computational cost. This study evaluates two hybrid models—Agglomerative K-Means and Divisive K-Means—where hierarchical clustering is used to initialize centroids, followed by K-Means refinement. This approach aims to reduce the limitations of random initialization while improving clustering stability and efficiency in large-scale data environments. Experiments on poverty data from Central Java Province show that hybrid methods accelerate K-Means convergence: Agglomerative K-Means reduced iterations to 2 (from 3 in standard K-Means), while Divisive K-Means converged in 1 iteration. Silhouette, Davies–Bouldin, and Calinski–Harabasz indices indicate that Agglomerative K-Means achieves the most compact and well-separated clusters, whereas Divisive K-Means performed worse than standard K-Means. Execution time measured only during the K-Means refinement phase shows that hybrids converge faster (Agglomerative: 2.04 ms; Divisive: 1.91 ms; K-Means: 116.68 ms), though this does not account for the hierarchical initialization cost. These findings provide practical insights into the trade-offs between clustering quality and computational efficiency when applying hybrid clustering methods. Overall, these results demonstrate that hybrid approaches can improve clustering stability and convergence efficiency, with Agglomerative K-Means providing the best balance between cluster quality and computational performance.