Research map / Label-efficient and robust learning / Recognition and long-tail

Metric-Agnostic Continual Learning for Sustainable Group Fairness

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 4 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 MacFRL, a label-sparse, metric-agnostic group-fair continual learner that reorders unlabeled covariate-shifted tasks by learned elastic-network depth and selectively augments a retained dataset with high-confidence instances.

Problem

Maintaining a group-fair classifier across a sequence of tasks with covariate shifts where only the first task is labeled, so that later tasks are unlabeled and distribution shifts cause negative transfer and unfair predictions.

Approach

Reorder the unlabeled task sequence by a task distance measured through learned layer-weight dynamics (Hedge BackPropagation over an over-complete elastic fair representation network, summarized by weighted entropy), perform adversarial domain adaptation with protected-feature obfuscation on the most similar task, and add high-margin predicted instances to the retained dataset.

Key results

• Across 8 datasets MacFRL outperforms five competitors on average by 12.7% accuracy, 42.8% demographic parity and 28.4% equalized odds (as reported by the authors).
• On Dutch census, MacFRL reaches 75.2% accuracy versus 52.6% for FaDL.
• On Bias-MNIST MacFRL achieves 92.9% accuracy and 14.1% equalized odds, tying the Skyline on accuracy-related comparisons.

Measured effects reported in the paper

Surprising findings

• Skyline (offline joint training with full labels) is outperformed by MacFRL on some fairness metrics, e.g. Dutch census equalized odds (0.115 vs 0.362 for Skyline).
• FaDL attains lowest demographic parity on Dutch census but at a large accuracy cost (52.6%).

Limitations

• Relies on a buffer of size k and selective augmentation whose choice balances reordering against sequential CL.
• Fairness is evaluated only with DP and EO differences; accuracy-fairness trade-off is sensitive to lambda1 (model collapses at lambda1=1).

Experimental setup

Datasets: Adult KDD Census-Income Bank marketing Dutch census Diabetes Law School Bias-MNIST CelebA

Benchmarks: Adult KDD Census-Income Bank marketing Dutch census Diabetes Law School Bias-MNIST CelebA

Metrics: average accuracy demographic parity difference equalized odds difference

Baselines: ULLC FaDL FaIRL UnFaIRL Skyline (joint multi-task learning)

Builds on: fair representation learning unsupervised domain adaptation with adversarial training Hedge BackPropagation FaIRL ULLC

Closest papers by meaning

• FRAPPÉ: A Group Fairness Framework for Post-Processing Everything (ICML 2024 · similarity 0.87)
• FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.86)
• Multigroup Fairness and Omniprediction: Separations and Equivalences (NeurIPS 2026 · similarity 0.86)
• Enhancing Group Fairness in Online Settings Using Oblique Decision Forests (ICLR 2024 · similarity 0.85)
• Stability and Multigroup Fairness in Ranking with Uncertain Predictions (ICML 2024 · similarity 0.85)
• FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (ICLR 2024 · similarity 0.85)