AUC Maximization Using Confidence Estimates for Biased Positive-Unlabeled Data
September 11, 2026
This method maximizes the area under the receiver operating characteristic curve when labeled positive data is biased and unlabeled data is abundant. It uses confidence scores—the probability of an instance being positive—to derive an AUC risk estimator that corrects for sample selection bias.
HOW THIS AFFECTS YOU
●
researcherYou can use this estimator to train more robust classifiers when your positive training samples do not represent the true population distribution.