Between-student differences explain a large fraction of CASED label variance, limiting what any clip-level visual model can learn¶

Claim · All claims
Evidence · 2 studies · 1 associational r1, 1 design r2 · q2
Currency · no dated study recorded · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Individual students differ substantially in mean engagement labels — some labeled engaged in over 90% of clips, others below 40% — so a large fraction of label variance reflects between-student differences rather than within-clip behavior. → student-1 q2 i? The binary label distribution is highly skewed (about 69% engaged, 31% not-engaged), so a majority-class classifier achieves approximately 40–45% macro F1 without learning discriminative features. → skew-1

Evidence¶

student-1¶

Alperen Kantarcı, Visvanathan Ramesh, and Gemma Roig. 2026. Mind the Student: Behavioral and Contextual Cues for Automated Engagement Prediction in Online Learning. ICMI '26, Napoli, Italy. https://doi.org/10.1145/3776574.3832485

q2 · i? · associational · r1

Dataset analysis of the CASED corpus (Section 2.1), examining per-student engagement rates. The authors draw the implication for clip-level visual models directly from this between-student variance.

"Several students are labeled engaged in over 90% of their clips while others fall below 40%. This hints that a large fraction of the label variance is explained by between-student differences rather than within-clip behavioral dynamics, which directly limits what any clip-level visual model can learn."

skew-1¶

Alperen Kantarcı, Visvanathan Ramesh, and Gemma Roig. 2026. Mind the Student: Behavioral and Contextual Cues for Automated Engagement Prediction in Online Learning. ICMI '26, Napoli, Italy. https://doi.org/10.1145/3776574.3832485

q2 · i? · design · r2

Dataset analysis of CASED label distributions (Section 2.1): approximately 69% of clips labeled engaged and 31% not-engaged, with continuous scores compressed (mean 3.7, median 4.0).

"A majority-class classifier therefore achieves approximately 40–45% macro F1 without learning any discriminative features."

Discussion¶

Related Claims¶

• The Bayesian context layer applies stronger posterior personalization to students with fewer training clips, correlating clip count with KL divergence from the prior (r=-0.748, p=0.001) — related
• Classroom-level ICCs for elementary test score gains range from 0.02 to 0.15 and school-level ICCs from 0.05 to 0.20, so classroom effects explain about 15 percent of gain-score variance — related
• Classroom and institutional differences account for little variance in engagement; nearly all variance is between students — related
• The integrated model explains substantial variance in student engagement (R² = 0.700) in GenAI-supported learning — related