A graph regularized overlapping community discovery framework with three-way decisions

Abstract

Community detection is essential for complex network analysis. Most existing approaches focus on hard community partitioning, and a few have investigated overlapping community structures, which are important but difficult to handle in practical applications. This paper presents a graph regularization-based framework for overlapping community detection, which integrates topological information and applies a theoretical three-way decision method to handle uncertain knowledge. The proposed models, $\text{GNMFO}\_\text{TW}$, $\text{GYNMFO}\_\text{TW}$, and $\text{GAEO}\_\text{TW}$, employ NMF, YNMF, and AEs with graph regularization terms for initial partitioning. The membership degrees of each node across different communities are then used for re-partitioning through three-way decisions. These models apply subspace clustering principles to incorporate basic network structure. To address the limitations caused by sparse network topology, the graph regularization terms encourage similar community membership among connected or nearby nodes, resulting in more coherent communities. In addition, three-way decisions, guided by node structural similarity, detect overlapping clusters and participating vertices. The proposed models not only identify community memberships but also reveal the overlapping community structures within networks. Empirical evaluations across both artificial and empirical networks indicate that our method outperforms existing advanced overlapping community detection techniques.