Research map / Learning theory and algorithms / Combinatorial and search optimization
Improving Robustness to Multiple Spurious Correlations by Multi-Objective Optimization
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 a debiased training method that groups data by bias-majority/minority pattern across multiple bias types and optimizes a softmax-weighted group loss with MGDA-inspired Pareto-stationarity regularization, plus a new real-image benchmark MultiCelebA.
Problem
Training unbiased accurate classifiers when a dataset contains multiple spurious correlations simultaneously, where mitigating one bias can exacerbate another and the bias-conflicting intersection is tiny.
Approach
Group samples by binary majority/minority indicators for each bias type (2^D groups each spanning all classes), then learn group-scaling weights via a loss that adds a Pareto-stationarity (MGDA-style gradient norm) term, which implicitly up-weights low-loss minority groups toward the minimax Pareto solution.
Key results
- On MultiCelebA two-bias setting, the method reaches 82.9% Unbiased and 82.6% CC, versus 63.8% Unbiased for ERM and 79.4% for Upweighting.
- On UrbanCars, CC accuracy is 87.6% (gap -4.2) versus 80.8% for the best prior bias-label method LISA.
- On Waterbirds the method achieves 91.8% Worst-group accuracy, and on BFFHQ 85.7% Unbiased, both best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MOO-based dynamic group-weight adjustment instead of the strongest prior baseline | debiased training · MultiCelebA two-bias | Unbiased accuracy: 80.9$_{\pm \text{0.2}}$ -> 82.9$_{\pm \text{0.2}}$ | Helped |
| use MOO-based dynamic group-weight adjustment instead of the strongest prior baseline | debiased training · MultiCelebA three-bias eval, two-bias training | Unbiased accuracy: 75.9$_{\pm \text{0.6}}$ -> 78.4$_{\pm \text{0.0}}$ | Helped |
| use MOO-based dynamic group-weight adjustment instead of GroupDRO | debiased training · Multi-Color MNIST | Unbiased accuracy: 78.7$_{\pm \text{2.7}}$ -> 80.0$_{\pm \text{2.0}}$ | Helped |
| use MOO-based dynamic group-weight adjustment instead of DFR | debiased training · Waterbirds | Worst accuracy: 90.2$_{\pm \text{0.8}}$ -> 91.8$_{\pm \text{0.3}}$ | Helped |
| use the proposed loss with dynamic group-weight adjustment rather than fixed uniform group weights (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 79.2$_{\pm \text{1.4}}$ -> 82.9$_{\pm \text{0.2}}$ | Helped |
| replace the proposed group-weight objective with weighted group-loss minimization alone (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 82.9$_{\pm \text{0.2}}$ -> 77.3$_{\pm \text{0.6}}$ | Hurt |
| replace the proposed group-weight objective with MGDA (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 82.9$_{\pm \text{0.2}}$ -> 82.3$_{\pm \text{0.4}}$ | Hurt |
| replace the proposed group-weight objective with GradNorm (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 82.9$_{\pm \text{0.2}}$ -> 81.9$_{\pm \text{0.6}}$ | Hurt |
| replace the proposed group-weight objective with MoCo (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 82.9$_{\pm \text{0.2}}$ -> 78.9$_{\pm \text{1.5}}$ | Hurt |
| group by the proposed binary guiding/conflicting labels g rather than bias attributes b (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 78.6$_{\pm \text{2.0}}$ -> 82.9$_{\pm \text{0.2}}$ | Helped |
| group by the proposed binary guiding/conflicting labels g rather than (b,t) (ablation) | Ours · MultiCelebA two-bias | Unbiased accuracy: 79.2$_{\pm \text{1.4}}$ -> 82.9$_{\pm \text{0.2}}$ | Helped |
| apply the proposed grouping strategy to GroupDRO instead of its original grouping (ablation) | GroupDRO · MultiCelebA two-bias | Unbiased accuracy: 78.4$_{\pm \text{0.7}}$ -> 78.8$_{\pm \text{1.4}}$ | Helped |
Surprising findings
- Learning group weights by minimizing the weighted group loss alone (without the Pareto term) performs worse than fixed uniform weights on most metrics.
- JTT, DebiAN, and DFR achieve Unbiased accuracy similar to or lower than plain ERM on MultiCelebA, and simple Upweighting beats most existing debiasing methods.
Limitations
- Requires bias attribute labels for training, which are expensive, especially in multi-bias settings.
- Minority groups in MultiCelebA are very small (CC has 150 training samples), and results depend on heavy subsampling to 95.3% bias-guiding.
- Assumes bias attributes are binary and known; authors note extending to unsupervised settings is future work.
Experimental setup
Datasets: MultiCelebA UrbanCars Multi-Color MNIST Waterbirds CelebA BFFHQ
Benchmarks: MultiCelebA UrbanCars Multi-Color MNIST Waterbirds CelebA BFFHQ
Metrics: Unbiased GG GC CG CC InDist Worst CCC Gap
Baselines: ERM LfF JTT DebiAN Upsampling Upweighting GroupDRO SUBG LISA DFR EIIL PGI
Builds on: MGDA GroupDRO LfF JTT
Closest papers by meaning
- Improving Subgroup Robustness via Data Selection (NeurIPS 2024 · similarity 0.85)
- SEBRA : Debiasing through Self-Guided Bias Ranking (ICLR 2025 · similarity 0.85)
- An Effective Theory of Bias Amplification (ICLR 2025 · similarity 0.84)
- The Impact of Coreset Selection on Spurious Correlations and Group Robustness (NeurIPS 2025 · similarity 0.84)
- Mitigating Spurious Correlations via Disagreement Probability (NeurIPS 2024 · similarity 0.84)
- Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random (NeurIPS 2024 · similarity 0.83)
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.