Edge embedding with global and local information fusion for community detection in complex networks

Community detection plays a crucial role in analyzing complex networks, including social and biological networks. However, current approaches often do not fully capture edge-level structural characteristics or effectively integrate multimodal information, resulting in unstable performance in networks with ambiguous boundaries and complex structures. We address these issues by proposing EdgeFusionNet, a novel edge-centric community detection framework comprising three core components. First, a parameter-free graph convolutional encoder extracts global node semantics from the network structure. Second, a multimodal edge representation module fuses local edge patterns with these global node semantics, generating discriminative edge embeddings. Third, a modularity-driven community aggregation algorithm iteratively refines community partitions through local node movement and community merging governed by a linearly decaying threshold, ultimately yielding globally optimized community structures. While recent research has begun to explore edge-centric formulations, EdgeFusionNet advances this paradigm by systematically addressing the key limitations of existing methods. Experimental results show that, compared with prevailing mainstream approaches across multiple real-world complex network datasets, the proposed method performs competitively or better while exhibiting consistent generalization capabilities and robustness.