Fold-local top-20 feature selection had mixed effects across model families, supporting the complete 166-dimensional representation as the primary analysis¶

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? Retaining the top 20 features improved both ordinal metrics for the linear and Random Forest models but worsened both for XGBoost and LightGBM; for AdaBoost, accuracy rose 0.9 points while macro-MAE and binary macro-F1 worsened slightly. → 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

Secondary sensitivity analysis refitting each fixed family with fold-local top-20 feature selection learned only from training participants (Table 4). The study reports a "family-dependent pattern" and describes the effects as mixed; all values are descriptive fold means.

"Feature-selection effects are therefore mixed rather than uniformly beneficial, supporting the complete 166-dimensional representation as the primary analysis."

Discussion¶

Related Claims¶

• 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 — related
• Model-family rankings depend on the prediction target: AdaBoost leads the five-level ordinal task while LightGBM holds the highest binary macro-F1 — related
• 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