Probabilistic-Gap-Driven Relabeling Method for Positive-Negative-Unlabeled Learning with Label Selection Bias

Label selection bias is common across different domains. In chronic disease management, many records lack gold standard tests which provide a definite label, but whether a record is labeled often depends on the patient’s health status. Our method introduces a probabilistic-gap–driven relabeling method that assigns pseudo-labels to unlabeled samples using probabilistic gaps rather than commonly used feature similarity. This algorithm employs a model-free design, which requires no parameter tuning. We integrate the relabeled data into a pseudo-likelihood framework with KMM-based reweighting method, mitigating selection bias and distributional shifts. Experiments on synthetic and real world medical datasets demonstrated that our approach effectively leverages mitigates selection bias and improves classification performance.