Fuzzy contrastive graph learning for overlapping community detection

Overlapping community detection in complex networks remains an important open problem, as real-world networks exhibit overlapping structures where nodes belong to multiple communities simultaneously. Existing methods either use hard assignments that fail to capture fuzzy memberships, or lack scalability and suffer from over-smoothing in deep architectures. We propose Fuzzy Contrastive Graph Learning (FCGL), a framework integrating fuzzy clustering with self-supervised contrastive learning through membership-aware graph augmentation, fuzzy contrastive loss with soft positive pairs, and multi-task optimization. Experiments on 8 benchmarks (up to 2.4M nodes) demonstrate that FCGL outperforms 9 state-of-the-art baselines by 11%–15% (relative improvement) in ONMI and F1-Score, with linear scalability and superior robustness. We analyze the convergence properties of the fuzzy contrastive objective and validate practical utility through case studies on social and biological networks. Our framework provides a principled approach bridging fuzzy set theory and contrastive representation learning for overlapping community detection. • Unified fuzzy clustering with contrastive learning for overlapping detection. • Fuzzy contrastive loss supports continuous weighted positive pairs. • Proven convergence, generalization bounds, and O(M) linear complexity. • Outperforms 9 baselines by 11–15% on ONMI and F1 across 8 datasets. • Membership-aware augmentation boosts performance and interpretability.