Att-MSR-GraphSAGE: An Attention-Guided Multi-Scale Residual Graph Neural Network for Human-Centric Stress-Factor Modeling

Human stress is a complex phenomenon influenced by behavioral, emotional, and contextual factors that interact in a non-linear manner. Traditional machine learning (ML) and deep learning (DL) models often consider these factors as independent variables, neglecting the inherent relational dependencies among them. This oversight leads to poor generalization and weak interpretability when modeling human-centric stress indicators. This research utilizes Graph Neural Networks (GNNs) to address these challenges, representing stress factors as nodes and their interrelations as edges, thereby facilitating topology-aware learning. The framework begins with a baseline GraphSAGE for localized aggregation and then extended to a Multi-Scale Residual GraphSAGE (MSR-GraphSAGE), which is capable of capturing hierarchical neighborhood information, with residual pathways that alleviate over-smoothing problems. Finally, a channel-wise attention fusion mechanism is adapted and integrated into the multi-scale GNN embeddings to dynamically recalibrate feature responses across propagation depths, thereby enhancing discriminative representation learning and inter-scale information integration for stress modeling. Experimental results on the Student Stress-Factor Dataset show that our model achieves an accuracy of 96.10% and F1-score of 94.21%. The proposed Att-MSR-GraphSAGE demonstrates superior performance compared to the evaluated baseline and residual GNN variants under the same experimental settings on the Student Stress-Factor dataset. To further assess generalizability, cross-dataset validation on the large-scale SWELL-KW database confirms the robustness and scalability of the framework. These results indicate that Att-MSR-GraphSAGE is a technically reliable, interpretable, and human-centric framework for intelligent stress-factor modeling.