Research map / Learning theory and algorithms / Prediction and decision losses
The Impact of Coreset Selection on Spurious Correlations and Group 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.
Empirical study of how coreset selection (5 difficulty scores, 5 selection policies, 10 spurious-correlation datasets) affects coreset bias level and downstream worst-group accuracy, finding that low bias does not reliably imply robustness, especially at small coreset sizes.
Problem
Coreset selection may amplify or mitigate dataset spurious correlations, yet its impact on downstream group robustness (worst-group accuracy) has not been systematically studied, especially without group labels.
Approach
A large-scale, multidimensional analysis that jointly measures coreset bias level and worst-group accuracy across learning-based (EL2N, Uncertainty, Forgetting) and embedding-based (SelfSup, SupProto) scores and multiple selection policies, with class balancing to isolate spurious-correlation effects.
Key results
- Learning-based scores (EL2N, Uncertainty) are substantially more informative than random for identifying bias-conflicting samples, whereas SelfSup and SupProto are near-random on Urbancars-C, Metashift, and Civilcomments.
- Difficult-selected EL2N coresets have lower bias levels than Easy or Random coresets but yield higher worst-group accuracy only when coreset size is sufficiently large (e.g., >10% on Waterbirds and Urbancars, >40% on Metashift and Civilcomments).
- At 5% and 10% coreset sizes, Difficult EL2N coresets suffer catastrophic average and worst-group accuracy drops (e.g., Civilcomments worst-group 5.0% at 10%), while Median and Stratified policies can recover robustness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EL2N Difficult selection instead of Random at a 40% coreset size | Downstream classifiers trained on spurious-correlation datasets · Waterbirds, Urbancars-B, Urbancars-A, MultiNLI, and Celeb-A hair | worst-group accuracy: 58.2 -> 74.6; 44.8 -> 52.0; 16.8 -> 23.2; 60.5 -> 46.0; 58.9 -> 38.3 | Hurt |
| Use EL2N Difficult selection instead of Random at a 40% coreset size | Downstream classifiers trained on spurious-correlation datasets · cMNIST, Metashift, Nico-spurious, and Civilcomments | worst-group accuracy: 62.0 -> 87.2; 73.8 -> 78.5; 44.0 -> 44.0; 77.4 -> 63.0 | Mixed |
| Prioritize difficult EL2N-scored samples rather than easy samples | Coreset selection on spurious-correlation datasets · 40% coreset size, ten datasets | worst-group accuracy | Mixed |
| Use difficult rather than easy selection at sufficiently large coreset sizes | EL2N-based coreset selection · Waterbirds, Urbancars, Metashift, and Civilcomments | worst-group accuracy | Helped |
| Use difficult selection in the very small coreset regime | EL2N-based coreset selection · Small coreset sizes across datasets | worst-group accuracy | Hurt |
| Select easy samples rather than difficult samples | EL2N-based coreset selection · Very small coresets, all datasets except cMNIST | average accuracy | Helped |
| Select easy samples rather than random samples | EL2N-based coreset selection · Very small coresets | worst-group accuracy | Hurt |
| Select difficult samples within each group rather than random samples within each group | Group-balanced EL2N selection · Waterbirds and Urbancars-C, very small coresets | worst-group accuracy | Hurt |
| Use Median or Stratified selection instead of Difficult selection | EL2N-based coreset selection · Very small coresets | worst-group accuracy | Helped |
| Exclude the highest-scoring 3% of samples from Difficult selection | EL2N Difficult selection (Difficult*) · Waterbirds and Urbancars-C at 2% and 5%; Metashift at 2%, 5%, and 10% | worst-group accuracy | Helped |
| Use Median selection instead of Difficult selection | EL2N-based coreset selection · Waterbirds and Urbancars-C at 10% | worst-group accuracy | Hurt |
| Use embedding-based scores rather than learning-based scores to select coresets | Coreset selection on real-world datasets · Urbancars-C, Metashift, and Civilcomments | bias level: 1.89-2.00; 1.87; 1.92-1.93; 1.92 | Helped |
Surprising findings
- Lower coreset bias level does not reliably guarantee higher downstream worst-group accuracy, particularly at small coreset sizes.
- Easy selection can yield high average accuracy at very small sizes while leaving worst-group accuracy well below random selection.
Limitations
- Limited to a finite set of 10 datasets and five scoring/selection methods, which do not cover all real-world spurious correlations.
- Group robustness is a coarse and limited notion of fairness.
Experimental setup
Datasets: cMNIST Waterbirds Urbancars-A Urbancars-B Urbancars-C Metashift Nico-spurious Civilcomments MultiNLI Celeb-A hair
Benchmarks: cMNIST Waterbirds Urbancars-A Urbancars-B Urbancars-C Metashift Nico-spurious Civilcomments MultiNLI Celeb-A hair
Metrics: worst-group accuracy average accuracy bias level average precision
Baselines: Random Random-Groupbalanced (R-Gbal) Difficult Easy Median Stratified
Builds on: EL2N forgetting score DRoP SupProto SelfSup
Closest papers by meaning
- Contributing Dimension Structure of Deep Feature for Coreset Selection (AAAI 2024 · similarity 0.87)
- Stable Coresets via Posterior Sampling: Aligning Induced and Full Loss Landscapes (NeurIPS 2025 · similarity 0.87)
- An Effective Theory of Bias Amplification (ICLR 2025 · similarity 0.86)
- $\mathbb{D}^2$ Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning (ICLR 2024 · similarity 0.86)
- Ameliorate Spurious Correlations in Dataset Condensation (ICML 2024 · similarity 0.86)
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 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.