Cluster-Perceptive Graph Contrastive Learning for Community Detection

Communities, revealing tightly-knit groups within networks, are essential to complex network analysis. While contrastive learning has been widely applied to community detection, existing studies often overlook latent community assignments information when constructing negative samples, resulting in many potential false negatives. To overcome this limitation, we propose Cluster-Perceptive Graph Contrastive Learning (CPGCL), which integrates the collaborative learning of intermediate representations and community assignments. Our approach utilizes a shared-parameter encoder to generate intermediate representations from two views. These representations are used to derive community probability distributions, which guide model training through contrastive loss between communities. Then by leveraging the generated sample community probability distributions to dynamically reweight the intermediate representations of sample pairs for contrastive learning, we mitigate the issue of potential false negatives and promote more discriminative intermediate representations. Additionally, high-confidence community assignments are used to iteratively refine the intermediate representations, further enhancing community detection performance. Experiments on three benchmark datasets validate the effectiveness and superiority of our approach. Code is available at https://github.com/l1tok/CPGCL.git.