ResearchPod Summary
Standard Positive-Unlabeled (PU) learning often assumes that labeled positive examples are selected completely at random (SCAR). However, in many real-world applications—such as medical diagnostics or recommendation systems—the labeling process is biased, meaning the probability of an example being labeled depends on its specific attributes. This paper investigates how to improve classification accuracy when this SCAR assumption is violated.
The authors propose the PUe (PU learning enhancement) framework, which adopts a causal inference perspective to address selection bias. The core of the approach is to estimate propensity scores—the probability that a positive example is labeled—and use these to reweight the training samples. Specifically, the authors introduce:
Experimental results on MNIST, CIFAR-10, and an Alzheimer’s disease dataset demonstrate that PUe consistently outperforms standard PU baselines under non-uniform label distributions. The improvements typically range from 1% to 5% across various metrics, including accuracy and AUC. The authors also show that the framework remains robust even when labels are evenly distributed, ensuring that the correction mechanism does not harm performance in unbiased scenarios.
By relaxing the restrictive SCAR assumption, PUe makes PU learning more applicable to real-world datasets where data collection is inherently biased. The integration with existing cost-sensitive methods provides a practical pathway for researchers to enhance their current models without needing to redesign their entire learning pipeline.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.