Research map / Label-efficient and robust learning / Recognition and long-tail
Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples
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 NSF, a task-independent debiasing method that identifies minority samples via relative distance to class centroids, estimates bias-invariant class means, learns a linear feature transformation to remove spurious features, and retrains the classifier on balanced groups, without bias labels.
Problem
Deep models learn spurious correlations between attributes and labels, and worst-group accuracy drops for minority groups when bias attributes are unknown and unlabeled.
Approach
Samples affected by spurious features disperse in feature space, so minority samples deviating from class centroids reveal bias; the unbiased class mean is estimated from majority/minority groups, and a learned channel-wise linear transformation aligns features to it before fine-tuning the classifier.
Key results
- Average WGA of 82.01% across four image/text benchmarks, versus 61.24% for ERM, with std of about 0.001 to 0.007 over 10 seeds.
- On Waterbirds, WGA rises from 72.60% (ERM) to 91.12% with mean accuracy 95.65%.
- On CheXpert, WGA reaches 70.21% versus 31.21% for ERM and 60.40% for JTT, with the method using no bias labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply NSF instead of ERM | Waterbirds · Waterbirds test | WGA: 72.60 -> 91.12 | Helped |
| apply NSF instead of ERM | CelebA · CelebA test | WGA: 47.20 -> 84.27 | Helped |
| apply NSF instead of ERM | MultiNLI · MultiNLI test | WGA: 68.11 -> 73.12 | Helped |
| apply NSF instead of ERM | CivilComments-WILDS · CivilComments-WILDS test | WGA: 57.06 -> 79.51 | Helped |
| apply NSF instead of ERM | CheXpert · CheXpert test | WGA: 31.21 -> 70.21 | Helped |
| finetune the classifier using transformed features without minority resampling (ablation) | NSF on Waterbirds · Waterbirds test | WGA: 72.60 -> 87.85 | Helped |
| use minority resampling with original features (ablation) | NSF on Waterbirds · Waterbirds test | WGA: 72.60 -> 86.92 | Helped |
| replace the identified Uk and Vk groups with randomly selected examples when learning t(x) (ablation) | NSF on Waterbirds · Waterbirds test | mean accuracy: 95.65 -> 90.60 | Hurt |
| replace the identified Uk and Vk groups with randomly selected examples when learning t(x) (ablation) | NSF on Waterbirds · Waterbirds test | WGA: 91.12 -> 73.99 | Hurt |
| discard feature channels using the lowest transformation coefficients instead of selecting channels randomly (ablation) | NSF · Waterbirds validation | WGA | Helped |
Surprising findings
- Finetuning the classifier on transformed features without minority resampling already gives a 15.25% WGA gain on Waterbirds, suggesting spurious-feature elimination alone is effective.
Limitations
- Relies on the strong spurious assumption that spurious features must be strong enough to be dominant; may not apply under weaker biases.
- Some classes cannot be estimated when the majority and minority groups are not both present (masked by o_i).
Experimental setup
Datasets: Waterbirds CelebA MultiNLI CivilComments-WILDS CheXpert
Benchmarks: Waterbirds CelebA MultiNLI CivilComments-WILDS CheXpert
Metrics: worst group accuracy mean accuracy
Baselines: ERM JTT MaskTune CNC AFR GroupDRO DFR uLA
Builds on: Deep Feature Reweighting (DFR) Just Train Twice (JTT) Group DRO Last-layer retraining Debiasing data model (Wang and Wang 2023)
Closest papers by meaning
- Severing Spurious Correlations with Data Pruning (ICLR 2025 · similarity 0.91)
- Complexity Matters: Feature Learning in the Presence of Spurious Correlations (ICML 2024 · similarity 0.89)
- Spurious Correlations in High Dimensional Regression: The Roles of Regularization, Simplicity Bias and Over-Parameterization (ICML 2025 · similarity 0.88)
- Mitigating Spurious Correlations via Disagreement Probability (NeurIPS 2024 · similarity 0.88)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.88)
- How Spurious Features are Memorized: Precise Analysis for Random and NTK Features (ICML 2024 · similarity 0.87)
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.