基于动态邻接融合与通道混合的图神经网络社团检测方法OA

Dynamic community detection has become a critical task in graph representation learning due to the dynamic evolu-tion of graph data in social networks and e-commerce platforms.Existing methods often model graph evolution using a unified time decay mechanism,struggling to characterize heterogeneous temporal behaviors.Moreover,they insufficiently model channel-wise feature interactions,thus limiting the balance between expressiveness and computational efficiency.To address these issues,this paper developed a novel dynamic graph learning framework,the temporal-channel graph attention network(TC-GAT).The TC-GAT framework integrated a dynamic adjacency fusion(DAF)module into a graph attention network(GAT)backbone.The DAF module achieved multi-stage adjacency information fusion through node-adaptive temporal weigh-ting,which effectively characterized diverse evolutionary behaviors.Furthermore,it introduced a graph channel mixer(GCM)to model deep interactions between channels in a lightweight manner,substantially enhancing node representation capabilities.Experimental results on multiple real-world dynamic graph datasets show that TC-GAT significantly outperforms mainstream models in key metrics such as accuracy,F,score,and AUC,while also demonstrating high training efficiency.These findings confirm that collaboratively modeling spatiotemporal evolution and channel interactions improves the overall performance of dy-namic graph analysis.