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

Learning Imbalanced Data with Beneficial Label Noise

ICML 2025 · Label-efficient and robust learning › Noisy labels · approach: Instance-dependent noise modeling

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 Label-Noise-based Re-balancing (LNR), which flips a subset of majority-class labels to the minority class using an asymmetric, instance-dependent noise model derived from classifier posteriors, to correct biased decision boundaries under class imbalance.

Problem

Class imbalance biases Bayesian and F1-optimal decision boundaries toward the minority class, and existing data-level methods (resampling, generative augmentation) cause information loss or generative errors.

Approach

Theoretically show that imbalance ratio shifts the accuracy-optimal boundary, then add asymmetric instance-dependent label noise that flips only majority samples near the boundary to the minority class, shifting the boundary back and raising minority-class F1/G-mean without adding or removing samples; the noise rate uses z-scored posteriors from any classifier, with a deferred schedule for multi-c

Key results

• On step-wise CIFAR-10, GCL+LNR reaches 72.22% AccMI vs 56.78% for GCL; LDAM+LNR gets 75.06% vs 66.41% for LDAM.
• On long-tailed CIFAR-100, MiSLAS+LNR achieves 48.52% overall accuracy, above MiSLAS+SelMix(10k) at 47.43%, with Few-shot accuracy 35.34% vs 25.10%.
• LNR achieves the highest average relative ranking across 32 KEEL datasets for KNN, CART and MLP in both F1 and G-mean; on imbalanced MNIST (IR=200), accuracy rose from 89.93% to 94.75%.

Measured effects reported in the paper

Surprising findings

• LNR flips only a small number of labels (e.g., 44 and 65 noisy labels on CIFAR-10/100-LT) yet yields large minority-class gains.
• On long-tailed CIFAR-100, LNR reduces Many-shot accuracy for GCL+LNR (67.16 to 57.11) while raising few-shot accuracy.

Limitations

• Theory covers binary classification with Gaussian-style examples; the multi-class extension is heuristic.
• Calibration (ECE) degrades for LNR on CIFAR-100 (9.55 vs 5.43 for MiSLAS) and mildly on CIFAR-10.
• Relies on a threshold t_flip and the flip-rate estimator's posteriors, which may be biased; tuned by grid search.

Experimental setup

Datasets: KEEL (32 binary datasets) CIFAR-10 CIFAR-100 MNIST (synthetic imbalanced illustration)

Benchmarks: KEEL 32 datasets CIFAR-10 step-wise CIFAR-100 step-wise CIFAR-10-LT CIFAR-100-LT

Metrics: F1 score G-mean AUC AccMI Overall accuracy Many/Medium/Few-shot accuracy ECE

Baselines: SMOTE ADASYN Borderline-SMOTE RUS OSS ClusterCentroids LDAM-DRW RSG GCL ReMix SelMix MiSLAS

Builds on: harmless label noise (Ahfock & McLachlan 2021; Cannings et al. 2020) LDAM-DRW MiSLAS RSG SelMix

Closest papers by meaning

• Toward Robustness in Multi-Label Classification: A Data Augmentation Strategy against Imbalance and Noise (AAAI 2024 · similarity 0.89)
• Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.87)
• Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · similarity 0.86)
• Improved Algorithm for Deep Active Learning under Imbalance via Optimal Separation (ICML 2025 · similarity 0.85)
• Synthetic Tabular Data Generation for Imbalanced Classification: The Surprising Effectiveness of an Overlap Class (AAAI 2025 · similarity 0.85)
• Imbalances in Neurosymbolic Learning: Characterization and Mitigating Strategies (NeurIPS 2025 · similarity 0.85)