Deep Fuzzy C-Means with Local Geometry Preservation

Deep clustering methods have shown promising results in unsupervised learning by jointly learning feature representations and cluster assignments. However, existing approaches often overlook the intrinsic geometric structure of data and rely on hard cluster assignments that may not capture the uncertainty in clustering decisions. In this paper, we propose Deep Fuzzy C-Means with Local Geometry Preservation (DFCM-LGP), a novel framework that integrates deep embedded clustering with fuzzy C-means initialization and manifold regularization. Our method preserves the local manifold structure during feature learning through a K-nearest neighbor graph-based constraint, while fuzzy membership enables soft cluster assignments that better reflect data uncertainty and inherently offers a more robust mechanism against noise and ambiguous data compared to hard assignments. We introduce a GPU-accelerated FCM implementation and a hybrid FCM-K-means initialization strategy to improve clustering quality. Extensive experiments on multiple datasets show that DFCM-LGP achieves improved performance compared to the baseline.