Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference
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 GIC, a group-inference method that trains a spurious-attribute classifier using a correlation (cross-entropy) term on training data and a KL spurious term contrasting training with comparison data, then plugs inferred groups into downstream invariant learners like Mixup and GroupDRO.
Problem
Group robustness under spurious correlations when group labels are unavailable; existing group-inference methods are inaccurate, leaving a large worst-group accuracy gap versus oracle group labels.
Approach
Infer spurious attribute labels with a classifier that is highly correlated with true labels on training data but whose label-conditional correlation differs from that on a (labeled or unlabeled) comparison dataset; the objective is cross-entropy minus a KL term between training and comparison conditional distributions, estimated with MINE, with a lower bound replacing the KL when comparison label
Key results
- GIC combined with Mixup (GIC_Cy-M) reaches 72.2% worst-group accuracy on CMNIST, 86.3% on Waterbirds, 89.4% on CelebA and 72.5% on CivilComments, outperforming group-label-free baselines such as JTT and EIIL.
- GIC's minority-group recall is consistently above 60% and exceeds 80% on CMNIST, CelebA and CivilComments.
- With Mixup on Waterbirds, GIC_Cy-M (86.3 worst-group) outperforms the oracle-label LISA baseline (89.2 worst-group is LISA's own result, so GIC is lower there); on CelebA GIC_Cy-M (89.4) is close to the oracle-label methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| infer groups with GIC using labeled comparison data and combine with Mixup rather than use JTT | worst-group classifier · CMNIST | worst-group accuracy (%): 67.3± \scriptstyle{5.1} -> 72.2± \scriptstyle{0.5} | Helped |
| infer groups with GIC using labeled comparison data and combine with Mixup rather than use CnC | worst-group classifier · Waterbirds | worst-group accuracy (%): 88.5± \scriptstyle{0.3} -> 86.3± \scriptstyle{0.1} | Hurt |
| infer groups with GIC using unlabeled comparison data and combine with Mixup rather than use CnC | worst-group classifier · CelebA | worst-group accuracy (%): 88.8± \scriptstyle{0.9} -> 89.5± \scriptstyle{0.0} | Helped |
| infer groups with GIC using labeled comparison data and combine with Mixup rather than use JTT | worst-group classifier · CivilComments | worst-group accuracy (%): 69.3 -> 72.5± \scriptstyle{0.3} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with GroupDRO | GroupDRO · Waterbirds | worst-group accuracy (%): 77.2± \scriptstyle{1.0} -> 80.2± \scriptstyle{0.1} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with GroupDRO | GroupDRO · CelebA | worst-group accuracy (%): 81.7± \scriptstyle{0.8} -> 82.1± \scriptstyle{0.3} | No clear effect |
| use GIC-inferred groups instead of EI-inferred groups with Subsample | Subsample · Waterbirds | worst-group accuracy (%): 81.9± \scriptstyle{1.4} -> 83.5± \scriptstyle{0.8} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with Subsample | Subsample · CelebA | worst-group accuracy (%): 82.8± \scriptstyle{0.5} -> 86.1± \scriptstyle{2.2} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with Upsample | Upsample · Waterbirds | worst-group accuracy (%): 81.3± \scriptstyle{0.7} -> 84.1± \scriptstyle{0.0} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with Upsample | Upsample · CelebA | worst-group accuracy (%): 84.8± \scriptstyle{0.2} -> 87.2± \scriptstyle{0.0} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with Mixup | Mixup · Waterbirds | worst-group accuracy (%): 85.7± \scriptstyle{0.4} -> 86.3± \scriptstyle{0.1} | Helped |
| use GIC-inferred groups instead of EI-inferred groups with Mixup | Mixup · CelebA | worst-group accuracy (%): 84.9± \scriptstyle{3.7} -> 89.4± \scriptstyle{0.2} | Helped |
Surprising findings
- Semantic consistency: GIC misclassifies short-haired women as men, and this misassignment yields Mixup pairs that better decouple spurious attributes than oracle groups do, so GIC+Mixup can beat oracle-label LISA on some tasks.
- Non-uniform comparison data constructed from the training set achieves worst-group accuracy comparable to using the validation set as comparison data.
Limitations
- Performance is weaker when group distributions of training and comparison data are very similar (e.g., CelebA), as noted by the authors.
- Gap to oracle group labels remains for several invariant methods, e.g., GroupDRO and EIIL variants.
- Relies on hyperparameter selection (gamma, epochs K) via a heuristic grid search tied to KL and CE loss trends.
Experimental setup
Datasets: CMNIST Waterbirds CelebA CivilComments-WILDS
Benchmarks: CMNIST Waterbirds CelebA CivilComments-WILDS
Metrics: average accuracy worst-group accuracy precision of minority group inference recall of minority group inference
Baselines: ERM GroupDRO LISA DFR SSA JTT CnC EIIL
Builds on: JTT EIIL LISA MINE GroupDRO
Closest papers by meaning
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.87)
- ExMap: Leveraging Explainability Heatmaps for Unsupervised Group Robustness to Spurious Correlations (CVPR 2024 · similarity 0.87)
- The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.86)
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.86)
- Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions (ICLR 2025 · similarity 0.86)
- Mitigating Spurious Correlations via Disagreement Probability (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.