Supplying learners' own engagement-definition codes as context raised exploratory prediction from 64.6% to 71.5% balanced 1-off accuracy in an eighteen-participant cohort¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? In an exploratory six-fold participant-independent analysis of 18 coded participants, an ordinal calibrator with definition–score interactions reached 71.5% balanced 1-off accuracy and 1.124 macro-MAE versus 64.6% and 1.216 for matched score calibration, a 6.9-point fold-average improvement. → Sidharth Anupkrishnan 2026

Evidence¶

Sidharth Anupkrishnan 2026¶

Sidharth Anupkrishnan, Itir Sayar, Jeongah Lee, Sri Harsha Musunuri, Guan-Ming Su, Madeline Endres, Ravi Karkar, Phuc Nguyen. (2026). E3Sense: Head-Confined Multimodal Sensing of Learner Engagement. arXiv. https://arxiv.org/abs/2609.26569 doi:10.48550/arxiv.2609.26569

q2 · i? · design · r2

Exploratory RQ2 calibration analysis: within six participant-grouped folds, four frozen per-modality engagement scores plus EDA site were compared with arms adding cognitive/behavioral/emotional definition codes and twelve interactions. The reported difference is descriptive, with no significance test.

"The interaction model reaches71 .5%balanced 1-off accuracy and1 .124 macro-MAE, compared with64.6%and1 .216for score calibration. The fold-average improvement is 6.9 percentage points in balanced 1-off accuracy."

Discussion¶

Related Claims¶

• Learners define engagement through cognitive, behavioral, and emotional criteria that align with the tripartite framework but differ across individuals — related
• A head-confined multimodal configuration (EEG, eye tracking, forehead EDA) predicts held-out learners' five-level engagement ratings at 75.0% balanced 1-off accuracy, 12.0 points above the mode baseline
• Fold-local top-20 feature selection had mixed effects across model families, supporting the complete 166-dimensional representation as the primary analysis
• Model-family rankings depend on the prediction target: AdaBoost leads the five-level ordinal task while LightGBM holds the highest binary macro-F1