AMON-Net: Integrating Graph Attention and Modularity Refinement for Community Detection in Complex Networks

Finding structural patterns in real-world networks requires community detection. Louvain and other modularity-based methods offer interpretable clustering, but they have scalability issues and resolution limitations. Strong embedding capabilities are provided by graph neural networks (GNNs), especially graph attention networks (GATs). However, because community-level optimization is generally lacking, these networks frequently fail to maintain structural coherence.We propose AMON-Net—an Attention-guided Modularity Optimization Network—for scalable and interpretable community detection. Three steps are included in AMON-Net: prototype-guided clustering in the embedding space, attention-weighted node embedding using GAT, and adaptive modularity refinement driven by modularity gain and attention uncertainty. Learned embeddings are further aligned with topological structures by a community-aware regularization term. Experiments on multiple benchmark datasets demonstrate that AMON-Net consistently outperforms both classical techniques and GNN-based baselines in terms of modularity, coverage, and clustering accuracy. These results demonstrate how well modularity-driven refinement and attention mechanisms work together for graph-based community detection.