Transfer-learned latent facial features support above-chance video-based mind wandering detection within a lab reading dataset, with LSTM reaching F1 = 0.459¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Within the lab reading dataset, models trained on latent AffectNet features outperform the random baseline to a significant extent, with F1 scores up to 0.459 (LSTM) and about 14% above chance. → Bühler 2025

Evidence¶

Bühler 2025¶

Bühler, B., Bozkir, E., Goldberg, P., Sümer, Ö., D'Mello, S., Gerjets, P., Trautwein, U., Kasneci, E. (2025). From the Lab to the Wild: Examining Generalizability of Video-based Mind Wandering Detection. Int J Artif Intell Educ 35, 823–857. https://doi.org/10.1007/s40593-024-00412-2

q2 · i? · design · r2

Person-independent 4-fold cross-validated evaluation on the lab dataset (Bosch & D'Mello, 2021; N = 135 students, 10-second video clips). Models on latent features outperformed the random baseline "to a significant extent"; Table 2 shows LSTM F1 = 0.459 and DNN AUC-PR = 0.398. The article notes F1 scores are driven by high recall with low precision.

"In the within-task prediction setting, models trained on latent features outperform the random baseline (i.e., F1 = 0.3 (Bosch & D'Mello, 2021) and AU C−PR = 0.3) to a significant extent."

Discussion¶

Related Claims¶

• Lab-trained mind wandering detectors generalize to in-the-wild lecture viewing across task, setting, and cultural group, with BiLSTM on latent features reaching F1 = 0.352 — related
• Mind wandering detection is more accurate for female than male learners, with the gender gap larger for temporal models — related
• Latent transfer-learned features yield better overall mind wandering detection than explicit OpenFace features, and temporal models add a small improvement, especially cross-dataset — a broader claim this one bears on
• Optimizing the classification threshold improves cross-dataset mind wandering detection, raising person-level correlation from 0.48 to 0.8 — related