Evaluation of student social networks using graph neural networks (GNN)

Abstract

Accurate assessment of student social networks is a central issue in the fields of educational big data and social computing. Existing methods largely rely on traditional statistical metrics or shallow network models, making it difficult to effectively capture the heterogeneity, dynamism, and higher-order dependencies of these relationships. Furthermore, these methods suffer from limitations such as feature coupling, poor interpretability, and a narrow scope of evaluation. To address this, this paper proposes a Multi-Channel Decoupled Attention Graph Neural Network (MCDA-GNN) for the comprehensive evaluation of student social networks. First, we construct a heterogeneous graph that integrates multidimensional behavioral and psychological features, distinguishing between strong and weak relationships as well as bidirectional and unidirectional interactions; Second, we design a topology-attribute dual-branch decoupling module that decomposes node representations into structural sharing and attribute complementarity components, eliminating feature redundancy through orthogonal constraints and mutual information maximization; Furthermore, we propose a dynamic multi-scale attention aggregation mechanism to adaptively capture higher-order dependencies and time-varying weights across different social circles, thereby enhancing the modeling capability of core-periphery structures; Finally, we establish a multi-dimensional evaluation framework encompassing structural connectivity, relationship stability, influence diffusion, and group cohesion, and design a graph-level contrastive learning loss to optimize model generalization. Experimental results on real student social datasets from three universities demonstrate that MCDA-GNN achieves significant improvements over mainstream models such as GCN, GAT, and GraphSAGE in tasks including relationship strength prediction, community partitioning quality, and influence assessment (with an average F1-score increase of 9.2% and an 8.5% improvement in NDCG@10), while maintaining stable performance under sparse-data and dynamic evolution scenarios. Ablation experiments validate the effectiveness of each innovative module, and case studies further demonstrate that the model can accurately identify key student nodes and fragile social links, providing a scientific basis for personalized educational interventions, campus community optimization, and mental health early warning.