Data-derived prototype profiles with GRPF-PPN for interpretable student stress prediction from psychometric tabular data

Student stress prediction requires models that are both accurate and interpretable. This study proposes GRPF-PPN, a Gated Residual Prototype-Fusion Psycho-Prototypical Network for classifying student stress from psychometric tabular data. The model combines feature-wise gating, residual representation learning, class-associated prototypes, and neural-prototype fusion to provide both discriminative predictions and case-comparison evidence. Across 10 repeated stratified splits of the main Student Stress dataset, GRPF-PPN achieved an accuracy of 0.929 ± 0.012, a Macro-F1 of 0.932 ± 0.013, a balanced accuracy of 0.930 ± 0.012, and an AUC of 0.973 ± 0.008. A target-overlap sensitivity analysis showed that a rule-based classifier achieved a Macro-F1 of 0.741 ± 0.029, while GRPF-PPN retained 0.881 ± 0.018 after target-proximal variables were removed. Strict external validation produced Macro-F1 scores of 0.858, 0.829, and 0.738 across three additional datasets, indicating heterogeneous transfer. Prototype purity was treated as internal class coherence rather than psychological validity. Prototype explanations were more stable than SHAP explanations under the reported perturbation protocol, although the two methods were interpreted as complementary. Overall, GRPF-PPN provides inspectable, data-derived prototype evidence for dataset-specific student stress classification.