Active positive-unlabeled (PU) learning addresses the common scenario where only positive and unlabeled examples are available, yet annotation budgets are tight. We introduce **A-PU-Net**, an end-to-end deep framework that: * **Selects informative negatives on-the-fly** via uncertainty-weighted margin sampling. * **Combines PU risk estimation** with a dynamic loss‐reweighting schedule for robustness to label noise. * **Balances exploration vs. exploitation** through a Bayesian stopping criterion that minimises redundant queries. Experiments on benchmark image and text datasets demonstrate up to **35 % fewer queries** at equivalent F1 score compared with baselines, highlighting the efficiency of our approach.Recommended citation: Kun Han, Weitong Chen, & Miao Xu. (2022). “Investigating Active Positive-Unlabeled Learning with Deep Networks.” *Proceedings of the 34ᵗʰ Australasian Joint Conference on Artificial Intelligence (AI 2021)*, pp. 607–618. https://doi.org/10.1007/978-3-030-97546-3_49
Download Paper | Download Slides
