|
Abstract
|
Community detection is a vital task in social network analysis, enabling the extraction of hidden structures and relationships. However, existing diffusion-based local community detection algorithms often depend on similarity-based scoring, which frequently failing to identify influential core nodes for expanding label. To address these shortcomings, we propose the local detecting and structuring communities (LDSC) method that integrates structural and relational insights with graph-based metrics and deep learning for refined community detection. LDSC uniquely stands out by combining Local Influence (LI) and Adaptive Absorbing Strength (AAS) metrics with GraphSAGE-based boundary refinement and adaptive community merging, tackling persistent challenges like scalability, boundary ambiguity, and structural cohesion unmet by prior methods. The method unfolds in four key phases: (1) Core Node Detection, employing a distinctive metric fusing LI and AAS to identify structurally significant nodes; (2) Label Diffusion, dynamically propagating labels from core nodes to neighbors for precise community formation; (3) Boundary Node Reassignment, using GraphSAGE to resolve ambiguities; and (4) Adaptive Community Merging, using an innovative local merging strategy to enhance cohesion. Evaluations on synthetic LFR benchmarks and real-world networks (e.g., Karate, Dolphins, DBLP, Amazon, LiveJournal, Orkut) demonstrate LDSC's superiority over baseline methods (e.g., LPA, CNM, WalkTrap, Louvain) and state-of-the-art approaches (e.g., Leiden, Infomap, LSMD, CLD_GE, FluidC, LCDR, LS), achieving perfect NMI/ARI (1.0) in Karate and Dolphins, top NMI in LiveJournal (0.92) and Orkut (0.65), average scores of 0.85 NMI and 0.75 ARI, and >15 % NMI improvement in large-scale networks like DBLP, showcasing unmatched accuracy, stability, and efficiency.
|