Subject-level demographic MAE-gap fairness regularization for multimodal attention models

Research Method · All research methods
Evidence · 2 claims (1 mixed, 1 against) · 1 study (1 design), q2 · 0 of 1 report an effect size · 2 claims rest on one study

Description

An in-processing fairness framework that adds a demographic disparity penalty to the training objective of a multimodal attention model. Per-subject MAEs are averaged within each demographic group, and "The fairness penalty was defined as the gap between the highest and lowest group-level subject MAE", weighted by a trade-off parameter λ added to the weighted MSE task loss. Regularization is introduced after a two-epoch warm-up, with subject- and group-balanced batch sampling so every represented subject contributes equally to the penalty. λ is selected on validation data subject to a constraint that mean validation MAE increases by no more than 0.02 relative to the unregularized baseline.

Accounts

• Subject-level demographic MAE-gap fairness regularization for multimodal attention models: An in-processing fairness framework that adds a demographic disparity penalty to the training objective of a multimodal attention model. Per-subject MAEs are averaged within each demographic group, and "The fairness penalty was defined as the gap between the highest and lowest group-level subject MAE", weighted by a trade-off parameter λ added to the weighted MSE task loss. Regularization is introduced after a two-epoch warm-up, with subject- and group-balanced batch sampling so every represented subject contributes equally to the penalty. λ is selected on validation data subject to a constraint that mean validation MAE increases by no more than 0.02 relative to the unregularized baseline. (Christoforos Fragkiadakis et al. (2026))

Claims

• Demographic MAE-gap regularization reduces targeted subgroup gaps on validation data, but these fairness gains do not consistently transfer to held-out test subjects [~M]
• Repeated subject-level split experiments confirm that validation-level fairness improvements from MAE-gap regularization do not correspond to consistent generalization to unseen subjects [-W]

Related Research Methods

-

Key Sources

• 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

Page metadata

Field	Value
type	research-method
id	subject-level-demographic-mae-gap-fairness-regularization-for-multimodal-attention-models
title	Subject-level demographic MAE-gap fairness regularization for multimodal attention models
description	An in-processing fairness framework that adds a demographic disparity penalty to the training objective of a multimodal attention model.
status	review
generated	by process:settle-candidates on 2026-10-11
trust	unverified — no one has confirmed this page's content

Frontmatter is the page's machine-readable record — see the Schema & Guide.