Research map / Learning theory and algorithms / Prediction and decision losses
Guiding Data Allocation for Robust Subpopulation Generalization
Zhaoying Pan, Yipei Wang, Shenyu Lu, Xiaoqian Wang
Read the paperCheck your idea against 137,217 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.
Studies whether fully balanced subgroup proportions are the unique optimal training configuration for robust subpopulation generalization, and argues that balance is costly and not always necessary.
Problem
Models achieve high average training accuracy but perform poorly under subpopulation shift; balancing subgroup proportions is a common data-centric fix but can be costly, infeasible, and not always necessary.
Approach
Questions whether balanced training data is the unique optimal configuration for robust subpopulation generalization, and instead studies how to guide data allocation across subgroups rather than enforcing full balance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| target Grad-derived configuration instead of balanced in under-sampling | Under-sampling · Waterbirds-syn (Example 2) | adjusted accuracy: 0.777 -> 0.901 | Improved |
| target Grad-derived configuration instead of balanced in over-sampling | Over-sampling · Waterbirds-syn (Example 2) | worst-group accuracy: 0.366 -> 0.541 | Improved |
| target Grad-derived configuration instead of balanced in under-sampling | Under-sampling · Waterbirds-syn (Example 2) | worst-group accuracy: 0.837 -> 0.848 | Improved |
| use balanced over-sampling instead of original data | Waterbirds-syn (Example 2) · Waterbirds-syn | worst-group accuracy: 0.376 -> 0.366 | No clear effect |
| use imbalanced Grad allocation for last-layer retraining stage two | DFR · Waterbirds/CelebA | worst-group accuracy: 89.30 -> 89.10 (Waterbirds); 87.49 -> 87.36 (CelebA) | No clear effect |
| use imbalanced Grad allocation for last-layer retraining stage two | AFR · CelebA | worst-group accuracy: 80.74 -> 83.84 | Improved |
| use imbalanced Grad allocation for last-layer retraining stage two | AFR · Waterbirds | worst-group accuracy: 86.90 -> 88.27 | Improved |
| use imbalanced Grad allocation for last-layer retraining stage two | GSR · CelebA | worst-group accuracy: 85.74 -> 86.67 | Improved |
| allocate data along Grad direction instead of balancing direction (ablation) | Data allocation · Waterbirds-syn, UrbanCars-syn, CelebA-subset, CivilComments-subset | adjusted accuracy | Improved |
| use non-balanced training configurations instead of balanced (ablation) | Subgroup data configuration · Gaussian theory and four datasets | adjusted accuracy: balanced 0.921 vs 90th percentile 0.918 | No clear effect |
Closest papers by meaning
- Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions (ICLR 2025 · similarity 0.87)
- Learning With Multi-Group Guarantees For Clusterable Subpopulations (ICML 2025 · similarity 0.87)
- Multigroup Robustness (ICML 2024 · similarity 0.87)
- Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift (ICML 2025 · similarity 0.86)
- Boosting Test Performance with Importance Sampling--a Subpopulation Perspective (AAAI 2025 · similarity 0.86)
- Group Distributionally Robust Optimization with Flexible Sample Queries (NeurIPS 2026 · 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.
