DI-CCNS: Directed community detection via co-clustering and node similarity with adaptive parameter optimization

Community detection in directed networks remains a challenging task due to directional asymmetry and complex topological relationships. While existing approaches like the DI-SIM spectral co-clustering framework are often limited by preset parameter sensitivity, insufficient modeling of directional information, and simplistic connectivity handling. This paper proposes a novel algorithm, DI-CCNS, based on direction-aware embedding and dynamic connectivity optimization, which address these limitations. First, we perform singular value decomposition (SVD) on the regularized Laplacian matrix through collaborative clustering to separate the node’s outgoing and incoming directional embedding vectors, precisely capturing the sending and receiving characteristics of nodes in directed interactions. Second, we introduce the Grey Wolf Optimization (GWO) algorithm to adaptively search for the optimal embedding dimensions and similarity threshold, eliminating reliance on manual parameter tuning. Furthermore, we propose three dynamic connectivity processing modes (DI-CCNS-SS/SD/WD) to optimize community merging strategies based on strongly/weakly connected components and directed neighbor relationships. Experiments on seven real-world datasets demonstrate superior performance over DI-SIM methods. The present algorithm not only achieves better results in the modularity of community structure but also shows stronger robustness in handling anomaly nodes, while requiring no preset community count and significantly enhancing robustness in complex directed scenarios.