Research map / Label-efficient and robust learning / Noisy labels

Optimal Transport Reweighting for Robust Learning under Spurious Correlations and Label Noise

NeurIPS 2026 · Label-efficient and robust learning › Noisy labels · approach: Noise-aware loss and pairwise methods · 0 citations (OpenAlex)

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 POTER, an optimal-transport reweighting method that derives per-sample importance from class-conditioned OT dual potentials between training data and a validation-built reference, enabling single-stage ERM robust to spurious correlations and label noise.

Problem

Worst-group accuracy under subpopulation shift from spurious correlations degrades further under label noise, since loss-based sample selection picks up mislabeled samples; group labels are only available on a limited validation set.

Approach

Derive sample weights from class-conditioned OT dual potentials between the noisy training empirical distribution and a reference distribution built from validation group annotations, then apply a KL-regularized closed-form exponential reweighting for a single ERM stage, avoiding loss-based selection and retraining.

Key results

• On Waterbirds at 30% symmetric label noise, POTER reaches 85.3% worst-group accuracy vs 68.9% for KNN-RAD.
• On CelebA at 30% symmetric label noise, POTER reaches 78.2% worst-group accuracy vs 70.2% for KNN-RAD.
• Under standard benchmarks POTER gets 72.5 (CMNIST), 90.9 (Waterbirds), 90.4 (CelebA) worst-group accuracy, best among validation-label-only methods on those three datasets (CivilComments 71.0 vs 72.9 for LEIA).

Surprising findings

• KNN-RAD falls from 88.1% to 7.5% worst-group accuracy on Waterbirds as subgroup-concentrated noise rises from 10% to 60%.
• Group DRO drops from 90.6% to 52.2% on Waterbirds at 30% symmetric label noise.

Limitations

• Depends on a representative, class-labeled validation set with group annotations; missing validation subgroups remain an open challenge.
• No finite-sample generalization guarantee for the weighted ERM classifier.
• On CivilComments POTER is slightly behind the best validation-label baseline (71.0 vs 72.9).

Experimental setup

Datasets: CMNIST Waterbirds CelebA CivilComments

Benchmarks: CMNIST Waterbirds CelebA CivilComments

Metrics: worst-group accuracy average accuracy

Baselines: ERM Group DRO LISA DFR PDE JTT CnC SSA AFR GIC LEIA KNN-RAD

Builds on: Sinkhorn algorithm Kantorovich duality Group DRO AFR KNN-RAD JTT

Closest papers by meaning

• Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions (ICLR 2025 · similarity 0.86)
• Retraining with Predicted Hard Labels Provably Increases Model Accuracy (ICML 2025 · similarity 0.85)
• Thumb on the Scale: Optimal Loss Weighting in Last Layer Retraining (NeurIPS 2025 · similarity 0.85)
• Learning from Noisy Labels via Self-Taught On-the-Fly Meta Loss Rescaling (AAAI 2025 · similarity 0.85)
• Self-Boost via Optimal Retraining: An Analysis via Approximate Message Passing (NeurIPS 2025 · similarity 0.85)
• Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining (ICLR 2025 · similarity 0.85)