Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Subgroup Robustness via Data Selection
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 D3M, which uses TRAK datamodels to find and remove training examples that most hurt worst-group accuracy, plus AUTO-D3M, which discovers pseudo-groups without annotations, to debias classifiers while removing far fewer examples than balancing.
Problem
Machine learning models fail on underrepresented subgroups; existing data-based debiasing such as dataset balancing needs training group annotations and removes large portions of data.
Approach
Approximate model predictions as linear functions of training data (datamodels estimated with TRAK), compute a group alignment score per training example weighted by group loss, and remove examples with negative alignment; AUTO-D3M derives pseudo-group labels from the top principal component of the TRAK matrix when no group labels exist.
Key results
- On CelebA-Age, D3M removes 2.4x fewer examples than data balancing (about 10K vs 30K) while improving worst-group accuracy.
- D3M reaches 75.6% WGA on CelebA-Age and 90.0% on CelebA-Blond, matching or beating methods with validation group labels; AUTO-D3M needs no group labels and gets 76.0% on CelebA-Age.
- Retraining CelebA-Age without five identified subpopulations raises WGA from 56.7% to 68.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply AUTO-D3M without train or validation group labels instead of ERM | Classifier training · CelebA-Age | worst-group accuracy (%): 56.7 -> 76.0 | Helped |
| Apply AUTO-D3M without train or validation group labels instead of ERM | Classifier training · CelebA-Blond | worst-group accuracy (%): 45.9 -> 83.8 | Helped |
| Apply AUTO-D3M without train or validation group labels instead of ERM | Classifier training · Waterbirds | worst-group accuracy (%): 57.9 -> 81.0 | Helped |
| Apply AUTO-D3M without train or validation group labels instead of ERM | Classifier training · MultiNLI | worst-group accuracy (%): 67.2 -> 75.0 | Helped |
| Use validation-group-guided D3M instead of JTT | Classifier training · CelebA-Age | worst-group accuracy (%): 61.0 -> 75.6 | Helped |
| Use validation-group-guided D3M instead of JTT | Classifier training · CelebA-Blond | worst-group accuracy (%): 81.6 -> 90.0 | Helped |
| Use validation-group-guided D3M instead of JTT | Classifier training · Waterbirds | worst-group accuracy (%): 63.6 -> 87.2 | Helped |
| Use validation-group-guided D3M instead of JTT | Classifier training · MultiNLI | worst-group accuracy (%): 72.6 -> 76.0 | Helped |
| Apply TRAK-based selection instead of ERM | ImageNet Red Wolf (Red Coat) classifier · ImageNet: Red Wolf (Red Coat) | worst-group accuracy (%): 22.62 -> 52.38 | Helped |
| Apply AUTO-TRAK selection instead of ERM | ImageNet Tench (Presence of human) classifier · ImageNet: Tench (Presence of human) | worst-group accuracy (%): 78.12 -> 80.00 | Helped |
Surprising findings
- Many examples in the most negatively aligned subpopulations of CelebA-Age contain labeling errors.
Limitations
- Requires validation group labels for D3M; AUTO-D3M relies on pseudo-labels and a hyperparameter lambda chosen per class.
- Removal count k is set by a heuristic that slightly overestimates the best k, or by searching on validation data.
- Theoretically no-information group robustness is unsolvable; results depend on coherent, data-isolated subpopulations.
Experimental setup
Datasets: CelebA-Age CelebA-Blond Waterbirds MultiNLI ImageNet
Benchmarks: CelebA-Age CelebA-Blond Waterbirds MultiNLI ImageNet
Metrics: worst-group accuracy balanced accuracy overall accuracy
Baselines: ERM RWG SUBG GroupDRO JTT DFR
Builds on: TRAK datamodeling Just Train Twice Deep Feature Reweighting GroupDRO
Closest papers by meaning
- Subgroups Matter for Robust Bias Mitigation (ICML 2025 · similarity 0.87)
- Discover and Mitigate Multiple Biased Subgroups in Image Classifiers (CVPR 2024 · similarity 0.85)
- Improving Robustness to Multiple Spurious Correlations by Multi-Objective Optimization (ICML 2024 · similarity 0.85)
- ExMap: Leveraging Explainability Heatmaps for Unsupervised Group Robustness to Spurious Correlations (CVPR 2024 · similarity 0.85)
- Rank-Constrained Adaptation for Reliable Real-World Performance (NeurIPS 2026 · similarity 0.84)
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.84)
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.