Contrastive graph clustering with Structure-Robust learning and stable prototype guidance

Contrastive graph clustering is an effective paradigm for unsupervised node representation learning, yet it still faces two challenges in high-dimensional attributed graphs. First, globally shared feature weights ignore node-wise variations in feature importance, reducing inter-cluster separability. Second, target distributions derived solely from current predictions lack temporal consistency, causing assignment fluctuations and prototype drift. To address these issues, we propose a unified framework termed Contrastive Graph Clustering with Structure-Robust Learning and Stable Prototype Guidance (SR-SPG). SR-SPG performs node-adaptive feature gating by integrating three priors: the neighborhood consistency prior, Fisher-discriminative prior, and adaptive learning prior, thereby enhancing feature separability. It also constructs temporally smoothed self-training targets via EMA-updated prototypes, thereby improving training stability. Extensive experiments on six benchmark datasets demonstrate the effectiveness and robustness of SR-SPG.