Yiming Fang, Heng Chang, Lanlan Qi

College of Electronic Engineering, National University of Defense Technology

Abstract：

With the increasing complexity and collaboration of cyberattacks, traditional single-point defense mechanisms are becoming less effective against large-scale organized threats. Community detection techniques, which identify tightly connected subgraph structures in complex networks, offer new insights for detecting attack clusters and blocking malicious propagation paths. However, existing graph neural network (GNN)-based community detection algorithms face challenges such as loss of global structural information and weakened interpretability in unsupervised scenarios. Traditional deep non-negative matrix factorization (DNMF) models, while effective in capturing global structural information and providing interpretable community partitioning results, struggle to effectively integrate node attribute information and have limited ability to capture local graph structural features. To address these issues, this paper proposes a Graph Neural Network-Enhanced Deep Non-negative Matrix Factorization algorithm for community detection (GEDNMF). GEDNMF leverages multi-layer graph convolutional networks to aggregate node attributes and high-order neighborhood information, generating low-dimensional embeddings. It introduces graph Laplacian regularization to maintain topological consistency, orthogonal constraints to reduce feature redundancy, and L2,1 sparse constraints to suppress noise interference, thereby achieving joint optimization of latent representations and enhancing the model’s generalization ability and stability. Experimental results on six benchmark datasets, including Cora and Citeseer, demonstrate that GEDNMF outperforms comparative algorithms such as NMF, VGAE, and DANMF in terms of clustering accuracy (ACC), normalized mutual information (NMI), and adjusted Rand index (ARI). For instance, on the Cora dataset, GEDNMF achieves an ACC of 60.08%, a 34.5% improvement over traditional NMF, with NMI and ARI values of 0.5096 and 0.3065, respectively, which are 28% and 8% higher than the second-best algorithm CDNMF. Ablation experiments further highlight the effectiveness of the proposed regularization mechanisms, showing that graph Laplacian regularization increases NMI by an average of 12.3%, orthogonal constraints boost NMI by an average of 16.7%, and L2,1 constraints improve accuracy by 0.75% in noisy environments.

Key Words：

community detection; graph neural networks; deep non-negative matrix factorization; regularization; unsupervised learning