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

Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification

NeurIPS 2026 · Label-efficient and robust learning › Noisy labels · approach: Instance-dependent noise modeling · 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.

Estimates the label-noise transition matrix column-by-column via one-sided selective classification (minimizing false discovery rate under a coverage constraint), with threshold and cost-sensitive algorithms that come with finite-sample bounds.

Problem

Estimating the label-noise transition matrix relies on fragile pointwise class-posterior estimation and existing estimators lack finite-sample guarantees; MPE methods with guarantees are mostly binary or intractable.

Approach

Frame each transition-matrix column as a one-sided selective classification problem (minimize false discovery rate for noisy label j subject to a minimum coverage constraint) and read off the empirical noisy-label distribution of accepted samples; this avoids class-posterior estimation, gives parametric-rate finite-sample bounds, and admits any binary classifier via threshold or cost-sensitive imp

Key results

• Proposed threshold and cost-sensitive algorithms achieve substantially lower transition-matrix MAE than the anchor-based method, whose error plateaus as n grows.
• Errors of the proposed algorithms decrease at a trend similar to a strong MLE oracle initialized at the ground-truth T.
• Theory gives error bounds of the form bias + O(1/sqrt(gamma n)), with bias controlled by false discovery rate suboptimality and coverage gamma.

Surprising findings

• The anchor-based method's error stagnates with more samples, due to its dependence on class-posterior estimation.

Limitations

• Scales poorly with the number of classes since a binary problem is solved per class.
• May struggle under class imbalance, affecting selection precision for minority classes.
• Bias term depends on R_gamma, which is nonzero when the anchor-point assumption fails.

Experimental setup

Datasets: Letter Satellite MNIST CIFAR-10

Benchmarks: Letter MNIST Satellite CIFAR-10

Metrics: mean absolute error of estimated transition matrix

Baselines: anchor-based estimator maximum likelihood oracle

Builds on: anchor-point estimation mixture proportion estimation selective classification with reject option

Closest papers by meaning

• Dual T: Reducing Estimation Error for Transition Matrix in Label-noise Learning (NeurIPS 2020 · similarity 0.93)
• Learning the Latent Causal Structure for Modeling Label Noise (NeurIPS 2024 · similarity 0.88)
• Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.87)
• Dirichlet-based Per-Sample Weighting by Transition Matrix for Noisy Label Learning (ICLR 2024 · similarity 0.87)
• Learning Causal Transition Matrix for Instance-dependent Label Noise (AAAI 2025 · similarity 0.87)
• Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.87)