An Improved Initialization Method for k-Means Clustering of Noisy Datasets Based on Rough Set Neighbourhood Model

Abstract

This study improves one of the initialization methods for the k-means clustering algorithm based on a rough set neighbourhood model to ‎enhance performance in noisy datasets. The method involves data normalization, obtaining a neighbourhood threshold based on the 0.25th ‎trimmed mean of pairwise Minkowski distances, calculating cohesion and coupling degrees of the neighbourhoods and between them re-‎spectively, and obtaining the initial cluster centres as the k points having maximum cohesion degrees with minimum coupling degrees among ‎themselves. The approach was evaluated on six datasets using Silhouette, Davies–Bouldin, Calinski–Harabasz, and Dunn–Hubert indices in ‎comparison with an existing method. Results showed that the improved method outperformed the existing method on noisy datasets, achieving higher Silhouette and Dunn–Hubert scores, and lower Davies–Bouldin values, with a slight reduction in Calinski–Harabasz index in ‎one of the datasets. On the non-noisy datasets, the two methods were at par in all four performance indices. With the improved performance, showing that the improved method enhanced the stability and robustness of k-means clustering in the presence of noisy data, it can ‎be recommended for clustering noisy datasets such as gene expression, image, and signal datasets‎.