Positive and unlabelled learning with generative adversarial networks

Abstract

The quantity of the data generated makes them difficult to process. In the case of supervised learning, labelling training examples may represent an especially tedious and costly task. In binary classification problems, examples are labelled as positive or negative. In this work, we assume that we have at our disposal a small number of positive and a larger number of unlabelled examples. Generative positive and unlabelled learning aims to generate labelled data, which constitutes a strategy to address this issue. Nonetheless, the generative approaches bring shortcomings, such as high computational cost, training instability and the inability to generate fully labelled datasets. We propose a novel generative approach based on an auxiliary classifier generative adversarial network. We integrate non-negative positive and unlabelled risk as an auxiliary loss to learn the distribution of positive and negative examples. We demonstrate the state-of-the-art performance on a common positive and unlabelled learning benchmark dataset. The results show that our approach achieves comparable performance as existing approaches despite its simple architecture, has high training stability and generates fully labelled data.