Demographic MAE-gap regularization reduces targeted subgroup gaps on validation data, but these fairness gains do not consistently transfer to held-out test subjects¶

Claim · All claims
Evidence · 1 study (2 entries) · causal 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? Gender-regularized training (λ=0.7) reduced the validation gender MAE gap from 0.02 to 0.005, but on the test set both the gender gap and worst-group MAE increased, and the gap was reduced in only 4 of 10 runs. → Christoforos Fragkiadakis 2026 q2 i? Age-regularized training (λ=0.8) reduced the validation age MAE gap from 0.105 to 0.093 and worst-group MAE from 0.352 to 0.345, but on the test set overall performance and demographic disparities worsened relative to the unregularized model. → Christoforos Fragkiadakis 2026 (2)

Evidence¶

Christoforos Fragkiadakis 2026¶

Christoforos Fragkiadakis, Seyed Sahand Mohammadi Ziabari, and Ali Mohammed Mansoor Alsahag. (2026). Fairness-Aware Multimodal Transformer Modeling for Real-Time Student Attention Estimation. https://arxiv.org/abs/2609.02232 doi:10.48550/arxiv.2609.02232

q2 · i? · causal · r2

Gender MAE-gap regularized Residual Fusion Transformer (λ=0.7) versus unregularized baseline, each across 10 seeds (Table 5). On test, the regularized model showed MAE 0.288 and gender gap 0.022 versus 0.283 and 0.015 unregularized.

"During validation, this configuration reduced the gender MAE gap from 0.02 to 0.005, while also slightly improving worst-group MAE, with a very small trade-off in overall predictive performance. However, these improvements did not generalize to the held-out test subjects."

Christoforos Fragkiadakis 2026 (2)¶

Christoforos Fragkiadakis, Seyed Sahand Mohammadi Ziabari, and Ali Mohammed Mansoor Alsahag. (2026). Fairness-Aware Multimodal Transformer Modeling for Real-Time Student Attention Estimation. https://arxiv.org/abs/2609.02232

q2 · i? · design · r2

Age MAE-gap regularized model (λ=0.8) versus unregularized baseline across 10 seeds (Table 6). On test, the regularized model showed MAE 0.292 and age gap 0.150 versus 0.283 and 0.126 unregularized.

"During validation, the age-regularized model reduced the age MAE gap from 0.105 to 0.093 and improved worst-group performance from 0.352 to 0.345, while incurring only a minor increase in overall MAE. However, these improvements did not generalize to unseen data."

Discussion¶

Related Claims¶

• The Residual Fusion Transformer shows the lowest worst-group MAE and smallest gender gap among baselines, while the Sensor GRU shows the smallest age gap alongside the weakest overall performance — related
• Repeated subject-level split experiments confirm that validation-level fairness improvements from MAE-gap regularization do not correspond to consistent generalization to unseen subjects — possibly the same claim (merge candidate)
• Facial visual dynamics provide the most informative cues for attention estimation: the Visual GRU outperforms both the train-mean baseline and the Sensor GRU — related