GSEC: Graph Structure Enhancement-Based Community detection using graph neural networks in attributed networks

Detecting communities in attributed networks is a fundamental yet challenging task. While Graph Neural Network (GNN)-based methods excel at this by integrating structural and attribute information, their performance is often constrained by the need to predefine the number of communities, K—a parameter typically unknown in real-world scenarios. Furthermore, their inability to flexibly leverage local pairwise prior knowledge further limits their effectiveness. To overcome these limitations, we introduce the Graph Structure Enhancement-based Community detection (GSEC) framework. GSEC integrates three key components: a hybrid-guided paradigm that co-optimizes the GNN architecture and learning objectives, a regular sampling mechanism to maintain graph sparsity, and an adaptive embedding refinement module to enhance node representations. This integrated design enables GSEC to maximize modularity and perform community detection without requiring a predefined K. Extensive experiments on five real-world complex networks demonstrate that GSEC outperforms eight state-of-the-art methods. It achieves exceptional performance with sparse supervision (e.g., 5%), exhibits strong robustness against noisy priors, and maintains high performance in fully unsupervised settings. The source code is available at: https://github.com/wy980125/GSEC.