Entropy-Guided Dissimilarity Measures for Enhancing Fuzzy C-Means Clustering Robustness

Fuzzy C-Means clustering is extensively employed in unsupervised learning tasks due to its capability to handle ambiguous and overlapping data structures. Nevertheless, classical Fuzzy C-Means is vulnerable to data noise and heavily influenced by the choice of distance metrics. This study systematically investigates the performance of entropy-guided dissimilarity measures combined with entropy-regularized objective functions to enhance clustering robustness. Specifically, we evaluate combinations of Euclidean distance and cosine dissimilarity with three entropy-guided variants, Entropy-Guided Dissimilarity, Normalized Entropy-Guided Dissimilarity, and Dual Entropy-Guided Dissimilarity, across synthetic datasets with controlled noise levels (0%, 5%, and 10%). Experimental results demonstrate the clear advantage of cosine-based clustering, exhibiting significantly higher Calinski-Harabasz Index scores (approximately 38.81–41.30) and greater stability under noisy conditions compared to Euclidean-based approaches (approximately 32.21–34.10). Interestingly, entropy-guided adjustments yielded minimal differentiation within distance metrics, yet slightly improved cluster stability under increased noise. Overall, this research underscores the robustness of cosine-based fuzzy clustering and provides practical insights into effective entropy-guided approaches for managing data uncertainty.