Predictive Adversarial Learning from Positive and Unlabeled Data

• DOI: 10.1609/aaai.v35i9.16953
• Publication Date: 2022-09-08T19:06:38Z
• Authors: Wenpeng Hu; Ran Le; Bing Liu; Feng Ji; Jinwen Ma; Dongyan Zhao; Rui Yan
• Keywords: Discriminator; Divergence (linguistics); Data set; Training set

This paper studies learning from positive and unlabeled examples, known as PU learning. It proposes a novel method called Predictive Adversarial Networks (PAN) based on GAN (Generative Networks). learns generator to generate data (e.g., images) fool discriminator which tries determine whether the generated belong (positive) training class. can be casted trying identify (not generate) likely instances set that determines identified are indeed positive. However, directly applying is problematic because focuses only data. The resulting will have high precision but low recall. We propose new objective function KL-divergence. Evaluation using both image text shows PAN outperforms state-of-the-art methods also direct adaptation of for