Mitigating Label Bias in Machine Learning: Fairness through Confident Learning

Yixuan Zhang, Boyu Li, Zenan Ling, Feng Zhou

Abstract

Discrimination can occur when the underlying unbiased labels are overwritten by an agent with potential bias, resulting in biased datasets that unfairly harm specific groups and cause classifiers to inherit these biases. In this paper, we demonstrate that despite only having access to the biased labels, it is possible to eliminate bias by filtering the fairest instances within the framework of confident learning. In the context of confident learning, low self-confidence usually indicates potential label errors; however, this is not always the case. Instances, particularly those from underrepresented groups, might exhibit low confidence scores for reasons other than labeling errors. To address this limitation, our approach employs truncation of the confidence score and extends the confidence interval of the probabilistic threshold. Additionally, we incorporate with co-teaching paradigm for providing a more robust and reliable selection of fair instances and effectively mitigating the adverse effects of biased labels. Through extensive experimentation and evaluation of various datasets, we demonstrate the efficacy of our approach in promoting fairness and reducing the impact of label bias in machine learning models.

Introduction

Machine learning models are fundamentally limited by the quality of their training data. A critical challenge arises when training data contains "label bias"—situations where ground-truth labels are overwritten by biased human agents, such as denying a loan to a qualified minority applicant. In these cases, models do not merely learn the intended task; they inherit and perpetuate existing prejudices.

Traditional solutions often attempt to mitigate this by adding fairness constraints to the loss function or adjusting model predictions post-hoc. However, these approaches frequently overlook the root cause: the biased data itself. While "Confident Learning" (CL) is a standard technique for identifying label errors by flagging instances where a model has low "self-confidence," it faces a significant hurdle in fairness contexts. Instances from underrepresented groups often receive low confidence scores not because they are mislabeled, but because the model lacks sufficient representative examples. Standard CL might inadvertently prune these fair but rare examples, potentially exacerbating bias.

Methodology

To address these challenges, we propose a data-centric approach that makes data selection fairness-aware through three key mechanisms:

  1. Truncated Confidence Scores: We utilize a robust M-estimator to truncate extremely high confidence scores. This prevents "overconfident" majority-group samples from disproportionately skewing the probabilistic thresholds used to identify errors, leading to a more stable selection process.
  2. Expanded Confidence Intervals: Rather than relying on a simple average threshold, we calculate the lower bound of confidence intervals. This acts as a "safety net" for instances with relatively lower confidence scores—common in minority groups—increasing their likelihood of being recognized as fair and included in the training set.
  3. Modified Co-teaching Paradigm: We employ a co-teaching framework where two identical networks are trained on different demographic subsets. These networks cross-validate each other to identify samples likely influenced by label bias versus "clean" samples. Unlike standard co-teaching, which targets hard-to-learn examples, our version focuses on identifying and removing samples in the "off-diagonals" of a joint distribution matrix, where observed labels are most likely to differ from true labels.
The proposed architecture utilizing a modified co-teaching paradigm with two networks (Net A and Net B) to select fair samples from biased datasets.

Experimental Evaluation

We evaluated our approach across five diverse datasets including Synthetic Data, Adult (income prediction), ProPublica COMPAS (recidivism), Credit Loan Data, and Law School Admissions.

Key Results

Our method consistently outperformed baselines, including standard Confident Learning (CL), LongReMix, and Group Peer Loss (GPL). Notable performance gains include:

  • Accuracy: On the "Adult" dataset with 20% symmetric bias, our method achieved a test error of 15.75%, compared to 22.61% for standard CL.
  • Fairness: On the "Law" dataset, the Difference of Equal Opportunity (DEO) was reduced to 0.05, a significant improvement over the 0.28 achieved by standard CL.
  • Stability: The approach demonstrated remarkable robustness. As shown in the figure below, even when bias levels were increased, the error rates remained relatively stable across different hyperparameter settings compared to baseline methods.
Error rates across different values of Ns for varying bias levels (0.2 and 0.4), demonstrating the stability of the proposed method.

Discussion and Limitations

While the proposed method effectively shifts the focus from algorithmic constraints to data quality, several areas for future work remain. The co-teaching element doubles the initial training cost, which may be a concern for very large-scale datasets. Additionally, the method's performance can be sensitive to hyperparameters like $N_s$ and variance ($\nu$), necessitating careful tuning via a validation set. Future research will explore extending this framework to multi-class sensitive attributes and intersecting identities to further broaden its applicability in real-world scenarios.

Summary updated

Abstract

Discrimination can occur when the underlying unbiased labels are overwritten by an agent with potential bias, resulting in biased datasets that unfairly harm specific groups and cause classifiers to inherit these biases. In this paper, we demonstrate that despite only having access to the biased labels, it is possible to eliminate bias by filtering the fairest instances within the framework of confident learning. In the context of confident learning, low self-confidence usually indicates potential label errors; however, this is not always the case. Instances, particularly those from underrepresented groups, might exhibit low confidence scores for reasons other than labeling errors. To address this limitation, our approach employs truncation of the confidence score and extends the confidence interval of the probabilistic threshold. Additionally, we incorporate with co-teaching paradigm for providing a more robust and reliable selection of fair instances and effectively mitigating the adverse effects of biased labels. Through extensive experimentation and evaluation of various datasets, we demonstrate the efficacy of our approach in promoting fairness and reducing the impact of label bias in machine learning models.
BibTeX
@inproceedings{zhang2024mitigating,
  author = {Yixuan Zhang and Boyu Li and Zenan Ling and Feng Zhou},
  title = {{Mitigating Label Bias in Machine Learning: Fairness through Confident Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=jXDmYHZ6Cc},
}
Figure