MotifStruCD: A two-stage approach for community detection via motif frequency analysis and structural optimization

Community detections considering motif-based higher-order features has demonstrated its effectiveness in distinguishing network organizations. However, traditional approaches are susceptible to feature drift resulting from the direct superposition of two-dimensional structures. In order to preserve nodes' structural characteristics while maximize information similarities of adjacent nodes, this study proposes a two-stage community detection framework (Motif-aware Structural Optimization for Community Detection, MotifStruCD) that integrates motif frequency analysis and structural optimization. In the first stage, node neighborhood distributions are extracted through a motif-based transfer probability matrix to preserve both higher-order and lower-order information, and then initial node feature vectors are generated by random walk. In the second stage, node features are first enhanced through graph attention networks, and structural optimization is further accomplished through contrastive learning. Experimental validation through clustering experiments on 19 datasets demonstrates that the MotifStruCD algorithm exhibits high effectiveness with a maximum increase up to 10.87% in the Q-value, and robustness against data noise or structural perturbations are further validated.