SCGAT: A self-supervised contrastive graph attention framework for community detection in attributed networks

The community detection in attributed networks plays a pivotal role in uncovering the hidden structures of complex systems. This process becomes challenging owing to the fact that the labeled data are scarce, and node attributes are noisy. This work proposes the Self-Supervised Contrastive Graph Attention (SCGAT) Framework that learns node embeddings without labels. The proposed method consists of the four components to extract the communities from the attributed network. The first one is the dual-view augmentation in which two complementary views of the same graph are constructed. The one view is an intra-community preserving (ICP) view with mild feature dropout to retain attribute coherence, and another is an inter-community disrupting (ICD) view with stronger dropout to blur community boundaries and promote invariance. The graph topology is kept fixed for both views. The second component is the attention-based encoding where each view is encoded by a two-layer Graph Attention Network. It assigns adaptive weights to neighbors and fuses structural and attribute signals into ℓ2-normalized embeddings. The another is pseudo-labelling in which ICP embeddings are clustered with K-Means to obtain community assignments and centres. The last one is multi-level contrastive learning. It minimizes a composite objective that (i) enforces node-wise consistency between ICP/ICD embeddings through an InfoNCE term, (ii) aligns ICD embeddings to the ICP community centers with a cross-entropy term, and (iii) adds a role-aware regularizer that stresses agreement for high-degree nodes, also called hubs. The model is optimized end-to-end in a self-supervised manner with fresh ICP/ICD views sampled each epoch to improve robustness. After training, communities are obtained by clustering the learned embeddings using K-means. Extensive experiments on four benchmark datasets (Cora, Citeseer, PubMed and Cornell) demonstrate that SCGAT consistently outperforms existing methods. On Cora, SCGAT achieves ACC = 0.8937 ± 0.0011, NMI = 0.7488 ± 0.0018, ARI = 0.7787 ± 0.0029 and F1 = 0.8844 ± 0.0014, representing improvements of +14−23% over the strongest baselines. Comparable gains are observed across Citeseer (12−18%), PubMed (9−19%) and Cornell (>10%), highlighting the scalability and robustness of SCGAT for community detection in real-world attributed graphs.