Improving Community Detection via Graph Neural Network-Based Edge Reweighting

Community detection methods such as Louvain and Leiden often operate on unweighted or heuristically weighted graphs and may therefore overlook latent relationships in noisy and structurally heterogeneous networks. To address this limitation, we propose a graph neural network (GNN)-based edge reweighting framework that learns adaptive edge weights from node representations while preserving the original optimization procedures of modularity-based community detection algorithms. The proposed framework employs a lightweight two-layer Graph Convolutional Network (GCN) trained with a contrastive learning objective using available node attributes or structural node descriptors when attributes are unavailable. The resulting node embeddings are used to estimate adaptive edge weights, producing a refined weighted graph prior to community detection. The proposed method is evaluated on eight benchmark datasets spanning citation, social, co-purchase, and synthetic networks, and is compared with both the original Louvain and Leiden algorithms and a Node۲Vec-based edge reweighting baseline. Across all datasets, the proposed approach consistently achieves higher modularity. For Louvain, it yields average improvements of ۷.۳۷% over the unweighted baseline and ۴.۷۱% over Node۲Vec-based weighting, while for Leiden, the corresponding improvements are ۷.۱۵% and ۴.۴۶%, respectively. Statistical analysis using the Wilcoxon signed-rank test with Bonferroni correction confirms that the modularity improvements over the unweighted baselines are statistically significant across all evaluated datasets. Additional experiments demonstrate improved robustness under edge perturbations, while ablation studies and hyperparameter sensitivity analysis validate the effectiveness and stability of the proposed framework. Overall, the proposed GNN-based edge reweighting framework provides an effective and robust preprocessing strategy for enhancing modularity-based community detection without modifying the underlying optimization algorithms.

کلیدواژه ها:

graph neural networks ، community detection ، Louvain Algorithm ، Leiden Algorithm ، Node۲Vec ، Edge Reweighting ، Contrastive Learning