Fairness-aware Model-agnostic Positive and Unlabeled Learning

Algorithmic bias is particularly severe in "positive and unlabeled" scenarios: The paper identifies that fairness problems are "usually more severe" in situations where only positive cases are labeled (e.g., diagnosed diseases in medical records) while negative cases remain unknown, yet this context has been largely overlooked in fairness research.

Real-world high-stakes decisions rely on incomplete data: Medical diagnosis, criminal risk assessment, and recommender systems typically only record confirmed positive outcomes, making it difficult to detect and correct for bias against specific social groups using traditional fairness approaches.

A new fairness framework (FairPUL) aims to equalize outcomes across populations: The proposed method targets equal true positive rates and false positive rates across different demographic groups, meaning the system should catch cases fairly and avoid false alarms equally for all populations.

The solution works with any machine learning model: By using a "model-agnostic" approach, the fairness framework can be applied after training any existing algorithm, making it practical for organizations already using different AI systems without requiring complete model redesign.

The method is mathematically proven to work: The framework demonstrates statistical consistency, meaning it reliably reduces both classification errors and unfairness simultaneously—addressing the common concern that fairness improvements come at the cost of accuracy.