GradPU: Positive-Unlabeled Learning via Gradient Penalty and Positive Upweighting

Positive-unlabeled learning is an essential problem in many real-world applications with only labeled positive and unlabeled data, especially when the negative samples are difficult to identify. Most existing positive-unlabeled methods will inevitably overfit class some extent due existence of unidentified samples. This paper first analyzes overfitting proposes bound generalization errors via Wasserstein distances. Based on that, we develop a simple yet effective method, GradPU, which consists two key ingredients: A gradient-based regularizer that penalizes gradient norms interpolated data region, improves class; An unnormalized upweighting mechanism assigns larger weights those hard, not-well-fitted less frequently labeled. It enforces training error each sample be small increases robustness labeling bias. We evaluate our proposed GradPU three datasets: MNIST, FashionMNIST, CIFAR10. The results demonstrate achieves state-of-the-art performance both unbiased biased scenarios.