Answers are grounded in this paper and cite their sources.
Communities in information networks often appear as cohesive subgraphs, reflecting strong relationships among entities. Given an information networkGand a query communityCq, similar community search (SCS) identifies a result communityCᵣ⊆ Gthat is similar toCqin terms of community-level features. SCS has broad applications in social marketing, online recommendation, and other domains. Supervised learning-based methods, such as neural subgraph matching (NSM) and graph alignment (GA), offer flexible learning and efficient inference, showing promising potential. However, they still have certain drawbacks. First, no existing method defines an effective metric to quantify the similarity betweenCqandCᵣat the community level. Second, GA emphasizes node-wise feature matching, but overlooks structure-wise matching, leading to matching failures when node features are missing. Third, the lack of high-quality human-annotated communities limits the effectiveness of supervised learning. To address these issues, we propose a unified metric for SCS that evaluates community similarity from multiple perspectives, including community cohesiveness, size, and density. We then propose SCS-GMN, an SCS solution based on the Graph Matching Network, which utilizes community similarity-guided self-supervised learning to enhance the cross-graph matching betweenGandCq, considering not only node-wise features but also structure-wise features. Moreover, we extend SCS-GMN to large graphs by introducing candidate subgraph generation and a graph-matching transformer. Empirical results over seven real-world datasets show that SCS-GMN efficiently returnsCᵣwith an average community similarity of 96.65% toCqwithin 64.5 ms, representing a 14.61% similarity improvement and a 2× inference speedup over suboptimal methods.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: