🤖 machine learning

A Data-Centric Framework for Detecting and Correcting Corrupted Labels

The paper introduces Relabeler, an end-to-end data-centric framework that leverages both local and global data relationships to detect and correct noisy labels, significantly outperforming state-of-the-art methods in label correction precision and downstream task performance.

Original authors: Ha-Linh Nguyen, Hong-Anh Nguyen, Minh-Duc La, Thu-Trang Nguyen, Son Nguyen, Hieu Dinh Vo

Published 2026-06-11

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Ha-Linh Nguyen, Hong-Anh Nguyen, Minh-Duc La, Thu-Trang Nguyen, Son Nguyen, Hieu Dinh Vo

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize different types of fruit. You show it thousands of pictures, but someone has carelessly stuck the wrong name tags on many of them. You've labeled a picture of an apple as a "banana," and a picture of a banana as an "apple." If you teach the robot using these messy instructions, it will learn the wrong things and fail when it tries to identify fruit in the real world.

This paper introduces a smart system called Relabeler that acts like a super-organized, hyper-observant librarian. Its job is to find those wrong name tags, figure out what the correct label should be, and fix them before the robot starts learning.

Here is how Relabeler works, broken down into simple steps:

1. The Problem: The "Noisy" Library

In the real world, data isn't perfect. Whether it's photos, text, or computer code, the labels (the names we give things) often contain errors. These errors are called "noise." If you try to learn from a noisy library, you end up with a confused student.

2. Step One: Finding the Suspicious Books (Detection)

Relabeler doesn't just guess; it uses two different ways to spot the wrong labels:

The "Local Neighborhood" Check: Imagine you are looking at a group of people standing together. If 9 out of 10 people are wearing red shirts, and one person is wearing a bright green shirt but is standing right in the middle of the red group, that person looks suspicious. Relabeler does this by looking at similar items (neighbors) in the data. If an item looks exactly like its neighbors but has a different label, it flags it as "suspicious."
The "Big Picture" Check: Sometimes, a person in a green shirt might actually belong in the red group (maybe they are just wearing a costume). To avoid false alarms, Relabeler takes a step back and looks at the whole library. It trains a smart model on the "clean" books it has already verified. Then, it asks this model: "Does this suspicious item fit the big picture?" If the model says, "Yes, this actually belongs here," the item is saved. If the model says, "No, this is definitely out of place," it stays on the "suspicious" list.

3. Step Two: Fixing the Wrong Tags (Correction)

Once Relabeler has a list of suspicious items, it doesn't just throw them away. Instead, it tries to fix the labels. This is the most unique part of the paper.

Think of this like a detective solving a mystery. The detective has two clues:

The Visual Clue: What does the item actually look like? (e.g., "This looks like an apple.")
The Mistake Pattern: How do people usually make mistakes? (e.g., "In this library, people often confuse apples with pears because they look similar.")

Relabeler combines these two clues using a mathematical method called Bayesian Inference. It asks: "Given that this looks like an apple, AND given that people in this dataset often mistake apples for pears, what is the most likely correct label?"

It calculates the probability and picks the best answer. It's not just guessing; it's making an educated repair based on how the data was messed up in the first place.

4. The Results: A Cleaner Library

The authors tested Relabeler on five different types of "libraries" (datasets), including images of cars, news articles, and computer code. They compared it to other top methods.

Better Accuracy: Relabeler was much better at fixing the labels correctly. In some cases, it improved the accuracy of the fixes by 58% compared to the best existing methods.
Fewer Mistakes Left Behind: After Relabeler finished its work, the remaining dataset had far fewer errors (up to 6% better performance in the final tasks).
Works on All Types of Noise: Whether the errors were random (like flipping a coin) or systematic (like confusing similar-looking things), Relabeler handled them all better than the competition.

The Bottom Line

The paper argues that instead of just teaching robots to ignore bad data (which is like telling a student to ignore the wrong answers), we should fix the data itself. Relabeler is a tool that finds the wrong answers, figures out what the right answers should be, and creates a clean, high-quality dataset. This allows any machine learning model trained on this new data to perform significantly better and more reliably.

Technical Summary: Relabeler – A Data-Centric Framework for Detecting and Correcting Corrupted Labels

Problem Statement
The performance of machine learning (ML) and deep learning (DL) models is fundamentally dependent on the quality of training data. In practice, large-scale datasets often contain "label noise"—instances assigned incorrect labels due to errors in crowd-sourcing or automated annotation. Reported noise rates in real-world datasets range from 8% to 38.5%. Such noise significantly degrades model accuracy and reliability. While existing approaches generally fall into two categories—model-centric (improving algorithm robustness) and data-centric (improving dataset quality)—most data-centric methods treat label detection and label correction as separate tasks. Furthermore, many existing solutions focus primarily on identifying noisy instances to discard or reweight them, rather than explicitly inferring and repairing the correct labels. This separation limits the ability to fully exploit the intrinsic relationships between input features, noisy labels, and clean labels within a unified framework.

Methodology: The Relabeler Framework
The authors propose Relabeler, an end-to-end data-centric framework designed to detect and correct corrupted labels through two integrated phases:

1. Corrupted Label Detection
This phase partitions a noisy dataset $\tilde{D}$ into a clean subset ( $D_{clean}$ ) and a suspicious subset ( $D_{suspicious}$ ) by leveraging both local and global instance relationships.

Local Verification: The framework employs the k-Nearest Neighbors (k-NN) algorithm to examine the consistency of an instance's label with its immediate neighbors in the feature space. Instances with labels consistent with the majority of their neighbors are tentatively classified as clean.
Global Verification: To address the limitations of local analysis (e.g., near class boundaries), Relabeler iteratively trains a classification model ( $M_{clean}$ ) on the currently identified clean instances. This model captures broader, global patterns. Instances in the suspicious set are re-evaluated; if their original labels align with $M_{clean}$ 's predictions, they are moved to $D_{clean}$ . This iterative process continues until no further clean instances can be identified.

2. Corrupted Label Correction
For instances remaining in $D_{suspicious}$ , Relabeler estimates the most probable clean label ( $y^*$ ) by maximizing the posterior probability $P(y | x, \tilde{y})$ , where $x$ is the input feature and $\tilde{y}$ is the observed noisy label. Using Bayesian inference, this is reformulated to combine two complementary factors:

Clean Label Prediction ( $P(y|x)$ ): A model $M_{clean}$ trained on $D_{clean}$ estimates the likelihood of a label based on input features.
Noisy Label Estimation ( $P(\tilde{y}|x, y)$ ): A model $M_{noisy}$ $M_{n o i sy}$ estimates the probability of observing a specific noisy label given the input and a candidate clean label.
- Training $M_{noisy}$ : Since ground-truth transition data is unavailable, Relabeler trains a "noise simulator" ( $M_{simulator}$ ) on $D_{suspicious}$ to learn the dataset's specific noise patterns. This simulator generates synthetic noisy labels for instances in $D_{clean}$ , creating a transition dataset used to train $M_{noisy}$ .
Inference: The final repaired label is determined by integrating predictions from both $M_{clean}$ and $M_{noisy}$ , ensuring the correction accounts for both semantic feature alignment and the statistical patterns of label corruption.

Key Contributions

Unified Framework: The introduction of Relabeler, an end-to-end system that jointly performs detection and correction, addressing the gap where detection and correction are typically treated separately.
Bayesian Integration: A novel approach to label correction that models both the clean label distribution and the noise transition probability, rather than relying solely on feature-based prediction or heuristic voting.
Iterative Refinement: A detection mechanism that iteratively refines the set of clean data to train a global verification model, improving the precision of suspicious instance identification.

Experimental Results
The authors evaluated Relabeler on five diverse datasets (CIFAR-10, CIFAR-100, Agnews, Clickbait, CodeXGLUE) covering image, text, and source code classification. The framework was compared against state-of-the-art baselines, specifically Docta (with voting and ranking variants).

Label Correction Precision: Relabeler consistently outperformed baselines, achieving up to a 58% improvement in label correction precision across various datasets and noise rates (20% and 40%). For example, on the Agnews dataset with 40% noise, Relabeler achieved 89.47% precision compared to 61.43% for Docta-R.
Error Rate Reduction: The framework significantly reduced the residual error rate in the repaired datasets. On CIFAR-10 with 40% noise, Relabeler reduced the error rate to 7.48%, whereas baselines retained rates above 23%.
Robustness: Relabeler demonstrated superior performance across different noise types, including symmetric, asymmetric, and instance-dependent noise. It maintained high precision and low error rates even under complex, mixed noise scenarios where baselines struggled.
Downstream Utility: Training downstream models (MLP classifiers) on datasets repaired by Relabeler yielded the highest accuracy and weighted F1 scores. On the Agnews dataset, Relabeler improved downstream accuracy to 75.50%, a ~6% relative gain over the original noisy dataset, outperforming both the original data and datasets repaired by Docta.
Component Analysis: Ablation studies confirmed that integrating the Noisy Label Estimator ( $M_{noisy}$ ) with the Clean Label Predictor ( $M_{clean}$ ) is critical, yielding up to a 50% relative improvement in precision compared to using $M_{clean}$ alone.

Significance and Claims
The paper claims that Relabeler addresses a critical gap in data-centric learning by moving beyond simple detection to explicit, principled label repair. By formulating label correction as a Bayesian inference problem that models both feature-label relationships and noise transition patterns, Relabeler provides a reusable, high-quality dataset that benefits any downstream learner. The authors assert that this approach is particularly practical for real-world applications where discarding data is not feasible, and where understanding the specific noise distribution is essential for accurate correction. The results suggest that Relabeler not only improves dataset quality but also serves as a robust complement to existing model-centric noise-robust learning methods.