Mitigating Label Bias via Decoupled Confident Learning

Label bias is a foundational problem that existing fairness methods overlook: The paper identifies that most algorithmic fairness approaches assume training labels are correct, but human-generated labels frequently encode societal biases—particularly in high-stakes domains like healthcare, hiring, and content moderation.

Biased labels can undermine fairness efforts: Even if an AI system is technically "fair," it may perpetuate discrimination if trained on mislabeled or biased data. This suggests that responsible AI requires scrutinizing data quality, not just model design.

DeCoLe offers a practical solution for identifying problematic labels: The proposed pruning method successfully identifies and removes biased labels from training data, demonstrating particular effectiveness in hate speech detection where labeling inconsistencies are known to be significant.

Hate speech detection reveals real-world label bias challenges: The paper's application to hate speech detection shows that label bias is not merely theoretical—it's a documented problem in systems deployed for content moderation, suggesting this issue affects platforms used by billions of people.

Addressing label bias improves model performance: By removing biased labels, DeCoLe outperforms competing approaches, indicating that cleaning biased training data is both an ethical imperative and a practical way to build better-performing AI systems.