PSPU: Enhanced Positive and Unlabeled Learning by Leveraging Pseudo Supervision
This addresses a specific bottleneck in PU learning for binary classification tasks, offering improvements in both balanced and imbalanced settings, though it appears incremental as it builds on existing PU methods.
The paper tackles the problem of overfitted risk estimation in Positive and Unlabeled (PU) learning by introducing PSPU, a pseudo-supervised framework that corrects model weights using confident samples and consistency loss, achieving significant performance gains on datasets like MNIST, CIFAR-10, and CIFAR-100.
Positive and Unlabeled (PU) learning, a binary classification model trained with only positive and unlabeled data, generally suffers from overfitted risk estimation due to inconsistent data distributions. To address this, we introduce a pseudo-supervised PU learning framework (PSPU), in which we train the PU model first, use it to gather confident samples for the pseudo supervision, and then apply these supervision to correct the PU model's weights by leveraging non-PU objectives. We also incorporate an additional consistency loss to mitigate noisy sample effects. Our PSPU outperforms recent PU learning methods significantly on MNIST, CIFAR-10, CIFAR-100 in both balanced and imbalanced settings, and enjoys competitive performance on MVTecAD for industrial anomaly detection.