Automated facial joy coding showed high sensitivity and specificity against human coding for both instructor and student videos¶

Claim · All claims
Evidence · 1 study (2 entries) · associational r1 · q3
Currency · newest study 2024 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q3 i? For instructor recordings, 93% of algorithm-identified joy expressions were also classified as joy by a human coder (sensitivity), and 97% of algorithm-coded non-joy frames were also coded non-joy by the human (specificity). → Frenzel 2024 q3 i? For student videos, algorithm sensitivity/specificity was 0.89/.88, and 84%/92% of human-coded joy/non-joy frames were congruently coded by the algorithm. → Frenzel 2024 (2)

Evidence¶

Frenzel 2024¶

Frenzel, A. C., Dindar, M., Pekrun, R., Reck, C., & Marx, A. K. G. (2024). Joy is reciprocally transmitted between teachers and students: Evidence on facial mimicry in the classroom. Learning and Instruction, 91, 101896. https://doi.org/10.1016/j.learninstruc.2024.101896

q3 · i? · associational · r1

Sensitivity and specificity analyses comparing AI-generated evidence scores against human judgements on systematically selected frames from each instructor and 13 students. The article reports 93% sensitivity and 97% specificity for instructor recordings, "assuming validity of the human coding".

"For the instructor recordings, 93% of the joy expressions identified by the algorithm were also classified as joy expressions by the human coder, thus demonstrating sensitivity of the system (assuming validity of the human coding)."

Frenzel 2024 (2)¶

Frenzel, A. C., Dindar, M., Pekrun, R., Reck, C., & Marx, A. K. G. (2024). Joy is reciprocally transmitted between teachers and students: Evidence on facial mimicry in the classroom. Learning and Instruction, 91, 101896. https://doi.org/10.1016/j.learninstruc.2024.101896

q3 · i? · associational · r1

Same validation procedure applied to student video frames. The article reports algorithm sensitivity/specificity of 0.89/.88 and human-to-algorithm congruence of 84%/92% for joy/non-joy frames, "assuming validity of the human coding".

"For the student videos, numbers were slightly lower, with an algorithm sensitivity/specificity of 0.89/.88 (assuming validity of the human coding)."

Discussion¶

Related Claims¶

• Instructors' and students' facial expressions of joy cross-recurred substantially above chance during a critical time window of about -2 to +3 seconds relative to instructors' expressions — related
• Facially expressed joy during class correlated positively with post-session self-reported session joy for both students and instructors, though the instructor correlation did not reach significance — related
• Instructor-led facial joy mimicry predominated: 58% of dyads were instructor-led and the critical lag window's midpoint was around +500 msec after the instructor's expression — related
• Instructors' post-session self-reported joy correlated positively with class-average instructor–student joy mimicry, but students' self-reported joy did not — related