Research map / Learning theory and algorithms / Prediction and decision losses
Multigroup Robustness
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.
Introduces multigroup robustness, requiring each subgroup's prediction shift under data corruption to scale only with corruption inside that subgroup, and gives a post-processing multiaccuracy boosting method that achieves it while preserving accuracy.
Problem
Standard learning algorithms (e.g., ERM) let corruption confined to some subpopulations shift predictions for unrelated subpopulations, so group-level harm is not bounded by the amount of corruption within each group.
Approach
Defines multigroup robustness as a data-aware bound on per-group mean prediction change proportional to corruption inside that group, shows multiaccuracy on the empirical distribution plus uniform convergence implies it, and proposes a post-processing multiaccuracy boosting step (reusing the full dataset each iteration, AdaBoost-style) that wraps any base learner while keeping its loss nearly unch
Key results
- Under label-change and data-addition poisoning on Folktables-Income, base logistic regression and MLP show growing multiaccuracy error on unattacked groups, while the post-processed predictors keep it consistently low.
- The post-processed predictor (Clf-PP) keeps accuracy similar to the base classifier across noise rates.
- Theory: the post-processing guarantees (3+m/n)epsilon multigroup robustness and 3epsilon multiaccuracy, and each update reduces empirical squared loss by at least epsilon^2, bounding iterations by 1/epsilon^2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| post-process the classifier with Multiaccuracy Boost using the training data | Logistic Regression on Folktables-Income · White male label changes; unaffected groups | multi-accuracy error | Helped |
| post-process the classifier with Multiaccuracy Boost using the training data | Logistic Regression on Folktables-Income · White male label changes; unaffected groups | accuracy | No clear effect |
| post-process the classifier with Multiaccuracy Boost using the training data | MLP on Folktables-Income · White male label changes; unaffected groups | multi-accuracy error | Helped |
| post-process the classifier with Multiaccuracy Boost using the training data | Logistic Regression on Folktables-Income · White female target; data added from White male group | multi-accuracy error | Helped |
| post-process the classifier with Multiaccuracy Boost using the training data | MLP on Folktables-Income · White female target; data added from White male group | multi-accuracy error | Helped |
| post-process the classifier with Multiaccuracy Boost using the training data | Logistic Regression and MLP on Folktables-Income · White female target; data added from White male group | accuracy | No clear effect |
Surprising findings
- Standard ERM-style classifiers can flip predictions for an entire group due to a tiny number of corrupted labels in a balanced distribution, illustrating that ERM is not multigroup robust even for simple subgroup classes.
Limitations
- Analysis limited to binary labels; multi-class setting left open.
- Guarantees use uniform convergence over a finite predictor class and require sufficiently large n.
- Multigroup robustness with respect to post-processed optimal actions (omniprediction) is not established.
Experimental setup
Datasets: Folktables-Income (ACS census, UCI Adult-like)
Benchmarks: Folktables-Income
Metrics: multiaccuracy error (MAE) per subgroup subgroup accuracy squared loss
Baselines: logistic regression two-layer MLP ERM
Builds on: multiaccuracy multicalibration AdaBoost
Closest papers by meaning
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.86)
- Learning With Multi-Group Guarantees For Clusterable Subpopulations (ICML 2025 · similarity 0.86)
- Multigroup Fairness and Omniprediction: Separations and Equivalences (NeurIPS 2026 · similarity 0.85)
- A Comprehensive View of Fairness through Distributional Stability (NeurIPS 2026 · similarity 0.85)
- Optimal Fair Learning Robust to Adversarial Distribution Shift (ICML 2025 · similarity 0.85)
- Like Oil and Water: Group Robustness Methods and Poisoning Defenses May Be at Odds (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.