Protocol-Aware Gradient Boosting for Context-Aware Student Engagement Detection

Abstract

We describe our submission to the ICMI 2026 Grand Challenge on Context-Aware Student Engagement Detection (CASED) for participant-independent binary classification of 30-second classroom clips (engaged vs. disengaged). We train a class-weighted CatBoost model on a 434-dimensional multimodal feature set spanning facial action units, gaze/head pose, eye/mouth aspect ratios, drowsiness cues, audio-event summaries, transcript TF–IDF SVD components, visual embeddings, and session-level context. Disengagement probabilities are converted to labels using a percentile threshold matched to the training base rate, improving robustness to score-distribution shifts. We further audit label structure and validation sensitivity, showing weak short-range temporal dependence and that early stopping/seed averaging/smoothing can materially affect GroupKFold-by-student OOF estimates. On the official held-out leaderboard, our system achieves F1macro = 0.52 and MCC = +0.04.