ORDINAL-EVIDENTIAL DYNAMIC HYPERGRAPH STATE-SPACE LEARNING FOR RELIABLE CHILD ENGAGEMENT RISK PREDICTION

The recognition of children's engagement and behavioural risk is critical to building a trusted AI-based approach to education and mental health support. However, current work shows that common engagement-recognition methods lack modelling of complex facial–temporal interactions, do not capture the ordering of engagement states, are prone to class imbalance, and often yield overconfident predictions in ambiguous behavioural settings. This study introduces OE-DHSSNet and the ER-MOBH optimisation strategy for uncertainty-aware and explainable engagement assessment. A selective state-space temporal encoder models long-range behavioural dependencies at low computational cost, and the framework establishes dynamic hypergraph relationships among behaviour-related, facial, and motion cues. An ordinal-evidential prediction head jointly estimates engagement levels and predictive uncertainty, while ER-MOBH jointly optimises predictive performance, calibration, and inference latency under constraints. The framework is built around the DAiSEE dataset, which provides four-level engagement labels, and also includes an EmotiW-based branch for out-of-distribution reliability analysis. In an illustrative, simulation-based evaluation designed to exercise the intended experimental protocol, the optimised OE-DHSSNet configuration reached an accuracy of 95.81%, a macro-F1 score of 94.45%, and a quadratic weighted kappa of 0.947, with an expected calibration error of 0.023 and an expected latency of approximately 32.6 ms per video sample. Explainability analysis further indicated that, for the 'Very High' engagement category, the eye/upper-face contribution (44%) was higher than for the 'Very Low' engagement condition (19%), whereas the head-pose/motion contribution was lower for 'Very High' (14%) than for 'Very Low' (42%). Evidential uncertainty correspondingly grew from 0.06 to 0.24, the expected direction for increasingly ambiguous behaviour. The proposed framework can thus help future educational-monitoring and early behavioural-risk-support AI systems, which must be trustworthy, by offering a unified approach to ordinal engagement recognition, uncertainty quantification and calibration, multi-objective optimisation, and interpretable behavioural assessment.