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¶

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? Across 10 seeds, the Residual Fusion Transformer achieved the lowest worst-group MAE for both age (0.328) and gender (0.289) and the lowest gender gap (0.015); the Sensor GRU achieved the smallest age gap (0.096) but with substantially lower predictive performance. → Christoforos Fragkiadakis 2026

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? · design · r2

Demographic subgroup evaluation across 10 seeds (Table 4): the fusion model also exhibited the lowest gender gap (0.015), while the Sensor GRU achieved the smallest age gap (0.096) and the Visual GRU the largest age gap (0.147).

"The Residual Fusion Transformer achieved the lowest worst-group MAE for both age (0.328) and gender (0.289), providing the most favorable balance between overall performance and worst-group error among the evaluated baselines."

Discussion¶

Related Claims¶

• A modality-aware reference model outperforms sensor-free, statistical, deep temporal, foundation-model, and LLM-based baselines for estimating minute-level attention difficulty from wearable sensing — related
• 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
• Demographic MAE-gap regularization reduces targeted subgroup gaps on validation data, but these fairness gains do not consistently transfer to held-out test subjects — related
• GRU is the lowest-performing model across both single-task and multi-task competency modeling, likely due to its two gating units versus LSTMs' three
• Chest-worn ECG provides the strongest cardiac-sensing performance for attention-difficulty prediction, and fusing all cardiac streams does not improve upon chest ECG alone — related
• Densely sampled visual frames and sensor representations offered no reliable predictive advantage over the 1 FPS setting in the higher-frequency experiment — 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 — related