A missing-aware temporal prototype hypergraph network with knowledge-guided explanation for joint student depressive-risk estimation and academic-stress prediction

Depressive-risk estimation and academic-stress prediction are important for scalable student mental-health support, because passive sensing can reveal behavioral signals linked to wellbeing in real-world campus settings. However, this task remains challenging because student data are heterogeneous, partially observed, temporally evolving, and often sparsely labeled, while mental-health risk is also shaped by shared contextual and academic factors rather than isolated individual behaviors. Existing approaches mainly include traditional feature-based machine learning, personalized sequence models, and graph-based depression detectors. Although these methods have advanced passive mental-health modeling, they often remain individual-centric, treat missingness as a preprocessing issue rather than a modeling signal, and provide limited evidence-grounded interpretability for actionable use. To address these limitations, we propose the Missing-aware Temporal Prototype Hypergraph Network (MTPHN), a unified framework for joint longitudinal student depressive-risk estimation and academic-stress prediction. MTPHN combines missing-aware multimodal encoding, multi-view dynamic hypergraph contrastive learning, prototype-guided risk scoring, and knowledge-guided post-hoc explanation with large language model verbalization. It captures temporal dynamics together with higher-order student relations, while preserving robustness under incomplete observations and generating semantically grounded reports. Experiments on three public real-world datasets, namely StudentLife, CES, and GLOBEM, show that MTPHN ranks first across all three benchmarks: on CES, MTPHN raises F1 from 0.874 to 0.893; on StudentLife from 0.842 to 0.859; and on GLOBEM from 0.853 to 0.870. Ablation, sensitivity, and explanation-faithfulness studies further verify that the proposed missing-aware fusion, dynamic hypergraph reasoning, prototype-based prediction, auxiliary stress supervision, and evidence-grounded explanation each contribute to robust prediction and faithful, structured interpretation under partial observations.