Evidence-Based Information Granulation for Three-Way FCM Clustering

Abstract

High-dimensional data often contain inherent ambiguity and complex local structures, which limit the performance of deep clustering methods. To address this, we propose a deep three-way FCM clustering method based on evidential information granulation. Our approach integrates contrastive learning into a deep FCM network to learn discriminative features. Using three-way decision theory, samples are divided into positive and boundary regions. A semi-ball neighborhood granulation method is then designed, and evidential theory is applied to fuse neighborhood trust degrees for precise sample reassignment. Experimental results on benchmark datasets show that our method outperforms state-of-the-art approaches in accuracy and normalized mutual information, demonstrating its effectiveness.