MOSAIC-Engage: Uncertainty-Aware Multimodal Deep Learning for Robust Student Engagement Estimation in Online Education
DOI:
https://doi.org/10.5281/zenodo.24388495Keywords:
component, formatting, style, styling, inseonline learning, student engagement estimation, multimodal deep learning, ordinal classification, reliability-aware fusion, selective prediction, uncertainty calibration.Abstract
Web-based teaching makes students less visible to instructors, but camera-facing behavior offers only a partial account of engagement. Looking away may signal distraction, reading, or reflection; equally, silent participation is not necessarily disengagement. Existing recognition studies also differ in labels, class balance, sensor inputs, and evaluation protocols, limiting the meaning of cross-paper accuracy comparisons. This article develops MOSAIC-Engage, a research framework for assessing ordinal engagement indicators from permitted video, audio, and optional interaction features. The proposed system uses independently measured signal-quality cues and availability masks to regulate multimodal fusion, followed by temporal encoding, coherent ordinal classification, and a calibrated option to withhold uncertain predictions. We distinguish two evaluation tracks: visual-only experiments on DAiSEE and genuinely multimodal experiments on CMOSE. A reproducible assessment plan specifies participant-disjoint splits, matched baselines, macro-F1 and ordinal metrics, modality-failure tests, participant-cluster uncertainty intervals, and end-to-end latency measurement. Published DAiSEE scores and CMOSE findings are discussed as secondary evidence, not as results of the proposed model. No MOSAIC-Engage training or deployment experiment is claimed. The principal contribution is a falsifiable architecture and an evaluation protocol addressing reliability, uncertainty, and educational validity without equating observable behavior with learning achievement.