Research map / Learning theory and algorithms / Fairness

FairMT: Fairness for Heterogeneous Multi-Task Learning

NeurIPS 2026 · Learning theory and algorithms › Fairness · approach: Fairness-constrained classification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FairMT, a fairness-aware multi-task learning framework that handles classification, detection, and regression under partial labels using asymmetric (one-sided) fairness disparities aggregated over tasks and optimized via a primal-dual scheme with a head-aware gradient proxy.

Problem

Fairness in multi-task learning with heterogeneous output types (classification, detection, regression) and partially missing task labels, where existing fairness methods are single-task, classification-oriented, and constrain only the shared encoder.

Approach

Asymmetric fairness gaps anchored to the best-performing group so only lagging groups are penalized, aggregated across tasks by a simplex water-filling/Frank-Wolfe scheme (AHFDA), combined with a primal-dual MTL objective and a head-aware Gram-matrix proxy for task weights.

Key results

• On CelebA, FairMT reaches 88.4% accuracy for the attribute task with Age EOD 3.4 and Gender EOD 6.4, versus 87.7% and 11.4/22.5 for MGDA.
• On AffectNet classification, FairMT reaches 60.6% accuracy with Age EOD 9.9, versus 59.5% and 23.4 for MGDA.
• On CiviComment regression, FairMT reaches CCC 73.7 with LGBTQ+ EOD 19.5, versus 73.4 and 27.4 for MGDA.

Limitations

• Code is promised only upon acceptance, not released.
• Fairness is evaluated on a limited set of sensitive attributes and datasets (CelebA, AffectNet, EmotioNet, CivilComments).
• Base architectures and training hyperparameters are deferred to the appendix, which is not included in the provided text.

Experimental setup

Datasets: CelebA AffectNet AffectNet-VA EmotioNet Civil Comments

Benchmarks: CelebA AffectNet AffectNet-VA EmotioNet Civil Comments

Metrics: Acc CCC EOD EO KS CSP

Baselines: MGDA GradNorm LNL UFaTE MGDA-Mean MGDA-UFaTE

Builds on: MGDA MGDA-UB GradNorm UFaTE LNL

Closest papers by meaning

• Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction (CVPR 2025 · similarity 0.91)
• Fair Resource Allocation in Multi-Task Learning (ICML 2024 · similarity 0.90)
• Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.87)
• Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.86)
• Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.86)
• FRAPPÉ: A Group Fairness Framework for Post-Processing Everything (ICML 2024 · similarity 0.86)