SHAP modality attributions aligned with each BARS dimension's behavioral focus, supporting interpretable feedback¶

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? SHAP analysis showed acoustic features dominated Voice & Talk (74.2% of predictive power) and Clarity, oculomotor features dominated Eye Contact, textual features dominated Topic and Content, and facial cues were primary for Nonverbal Expressiveness. → Hung-Yue Suen and Kuo-En Hung 2026

Evidence¶

Hung-Yue Suen and Kuo-En Hung 2026¶

Hung-Yue Suen and Kuo-En Hung. (2026). An Interpretable Closed-Loop Intelligent Tutoring System for Multimodal Affective Feedback in Asynchronous Presentation Training. IEEE Transactions on Learning Technologies. https://doi.org/10.1109/TLT.2026.3693864

q2 · i? · design · r2

SHAP attributions aggregated to the modality level and normalized within each of the seven models (Table III); modality importance was not proportional to feature count despite the 3,780-dimensional facial feature set.

"For Voice & Talk, the 203 acoustic features collectively contributed 74.2% of predictive power, indicating that fewer, highly informative features can outweigh larger but more redundant feature sets."

Discussion¶

Related Claims¶

• Adult learners using the closed-loop ITS improved significantly across all seven BARS presentation dimensions over a 30-day practice window (Cohen's d = 0.39–0.90) — related
• Multimodal XGBoost models achieved rubric-aligned scoring comparable to expert ratings, and no single modality approached multimodal performance — related
• Facial features of negative emotions predict low rapport, but additional modalities added no predictive power — related