Improving Clustering Robustness through Fuzzy Ensemble of K-Means and Mean Shift

Abstract

Ensemble clustering has emerged as a powerful strategy to improve the robustness and accuracy of unsupervised learning, particularly when individual algorithms struggle with noisy, heterogeneous, or high-dimensional data. This study introduces a fuzzy-based ensemble approach that integrates the complementary strengths of K-Means and Mean Shift clustering, followed by fuzzy membership assignment for data points that remain ambiguous. The inclusion of fuzzy logic provides a flexible mechanism to resolve uncertainty, ensuring that overlapping or irregularly shaped clusters are effectively managed. Experiments were conducted on three benchmark datasets-Weather History, Weather Prediction, and Dry Bean-using evaluation metrics such as the Silhouette Score and Davies–Bouldin Index. Results show that the proposed ensemble achieves consistent improvements over traditional clustering methods, with significant reductions in Davies–Bouldin Index and higher Silhouette Scores across datasets. These findings highlight the practical potential of the method for complex real-world applications and contribute to advancing ensemble clustering methodologies.