Research map / Learning theory and algorithms / Prediction and decision losses
Modeling the Second Player in Distributionally Robust Optimization
Paul Michel, Tatsunori Hashimoto, Graham Neubig
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use P-DRO instead of ERM without group information | BiasedSST · BiasedSST test | robust accuracy: 2.15 ± 0.97 -> 34.98 ± 9.39 | Improved |
| Use P-DRO instead of NonParam | BiasedSST · BiasedSST test | robust accuracy: 28.11 ± 2.16 -> 34.98 ± 9.39 | Improved |
| Use P-DRO instead of ERM without group information | DWMW17 · DWMW17 test, no group information | robust accuracy: 53.19 ± 1.70 -> 69.06 ± 1.70 | Improved |
| Use P-DRO instead of NonParam without group information | DWMW17 · DWMW17 test, no group information | robust accuracy: 54.13 ± 1.14 -> 69.06 ± 1.70 | Improved |
| Use P-DRO instead of ERM without group information | FDCL18 · FDCL18 test, no group information | robust accuracy: 19.57 ± 7.00 -> 30.25 ± 10.13 | Improved |
| Use P-DRO instead of ERM when validation group annotations are available | DWMW17 · DWMW17 test, group information on validation set | robust accuracy: 53.15 ± 0.87 -> 63.05 ± 4.25 | Improved |
| Use P-DRO instead of ERM when validation group annotations are available | FDCL18 · FDCL18 test, group information on validation set | robust accuracy: 34.07 ± 3.20 -> 47.61 ± 4.53 | Improved |
| Add the KL constraint to Minmax stopping (ablation) | P-DRO · BiasedSST validation, stopping criterion | robust accuracy: 25.22 ± 13.01 -> 31.30 ± 10.07 | Improved |
| Use Greedy-Minmax stopping with the KL constraint (ablation) | P-DRO · BiasedSST validation, stopping criterion | robust accuracy: 31.30 ± 10.07 -> 32.17 ± ± 11.20 | Improved |
| Replace Average stopping with Minmax stopping (ablation) | P-DRO · BiasedSST validation, stopping criterion | robust accuracy: 0.00 ± 0.00 -> 25.22 ± 13.01 | Improved |
| Use KL-constrained Minmax rather than Average for hyperparameter selection (ablation) | P-DRO · BiasedSST validation, hyperparameter selection | robust accuracy: 31.03 ± 12.16 -> 35.65 ± 11.47 | Improved |
| Add the approximate KL constraint to bare P-DRO (ablation) | P-DRO toy task · BiasedSST toy task | robust accuracy: 49.85 ± 0.10 -> 58.41 ± 9.25 | Improved |
Closest papers by meaning
- Large-Scale Non-convex Stochastic Constrained Distributionally Robust Optimization (AAAI 2024 · similarity 0.90)
- Distributionally Robust Optimization with Bias and Variance Reduction (ICLR 2024 · similarity 0.90)
- Differentiable Distributionally Robust Optimization Layers (ICML 2024 · similarity 0.88)
- Outlier-Robust Distributionally Robust Optimization via Unbalanced Optimal Transport (NeurIPS 2024 · similarity 0.88)
- Wasserstein Distributionally Robust Optimization through the Lens of Structural Causal Models and Individual Fairness (NeurIPS 2024 · similarity 0.88)
- Structure-informed Risk Minimization for Robust Ensemble Learning (ICML 2025 · similarity 0.88)
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.
