Robust multi-view fuzzy clustering with exponential transformation and automatic view weighting
Zhe Liu, Haoye Qiu, Muhammet Deveci, Sukumar Letchmunan, Luis M. Martínez · Knowledge-Based Systems · 2025
Multi-view fuzzy clustering has gained widespread attention due to its unique capability to handle uncertainty through flexible membership assignment, allowing samples to belong to multiple clusters with varying supports, thereby providing a comprehensive understanding of multi-view data. This capability is particularly relevant to knowledge-driven systems that require interpretable integration of multi-view data. However, existing multi-view fuzzy clustering algorithms often struggle with handling noise and incorporating flexible weighting strategies for different views effectively. To address these challenges, this paper proposes four robust multi-view fuzzy clustering algorithms (RMFC-ET-VS, RMFC-ET-VP, RMFC-ET-MS, RMFC-ET-MP), which leverage an exponential transformation of Euclidean distance to effectively mitigate the impact of noise and outliers in the data, thereby enhancing clustering stability. Moreover, we introduce vector-based and matrix-based view weighting strategies, employing sum-to-1 and product-to-1 constraints to ensure that the most informative views contribute more effectively during clustering. The proposed algorithms offer a dual emphasis on robust distance metrics and adaptable view weighting, resulting in more accurate and resilient clustering outcomes. Extensive experiments on multiple real-world datasets demonstrate that the proposed algorithms significantly outperform existing multi-view clustering algorithms, both in terms of clustering performance and robustness. • This paper proposes four robust multi-view fuzzy clustering leveraging exponential transformation. • It introduces vector-based and matrix-based view weighting with sum-to-1 and product-to-1 constraints. • Flexible view-level and cluster-level weight adjustments improve adaptability and accuracy. • Experiments show that the proposed algorithms outperform the existing algorithms.