Fuzzy Clustering Means Integrated Multi-Level Learning for Contrastive Multi-View Clustering

Multi-view clustering (MVC) effectively mines complementary information from diverse modalities, yielding more accurate and robust results than single-view methods. Recent approaches address key challenges in traditional MVC, target conflict and private information interference, through multi-level feature decoupling and fusion-free mechanisms. However, during the transition between contrastive learning and fine-tuning stages, existing methods often rely on rigid pseudo-label generation techniques such as hard labels. This rigidity overlooks the inherent fuzziness of category boundaries in real-world data, leading to issues including view discrepancy, which causes ineffective cross-view semantic alignment, and manifold blindness, which denotes failure to capture complex data structures. To overcome these limitations, this paper proposes FCM-MFL-MVC, a novel multi-view feature learning framework that integrates Fuzzy C-Means (FCM) clustering within a contrastiveMVCparadigm for pseudo-label guided MVC. Our key innovation lies in generating probabilistic pseudo-labels via FCM, naturally modeling cluster assignment uncertainty and mitigating erroneous label propagation. This core idea is supported by a dynamic learning rate scheduler to stabilize optimization across stages and a balanced dual-objective loss to enhance clustering quality while suppressing noise-induced overfitting. Extensive experiments on high-dimensional datasets including CCV, Caltech, ACM, MNIST-test and EYB-fc demonstrate that FCM-MFL-MVC significantly outperforms state-of-the-art methods, achieving marked improvements in both clustering performance measured by ACC and NMI and in model robustness.