Conditional generative positive and unlabeled learning

doi:10.1016/j.eswa.2023.120046

2023-08-01

Aleš Papič *

Igor Kononenko

Zoran Bosnić

DOI：10.1016/j.eswa.2023.120046

原文链接

查看原文

摘要

En 中文

The quantity of data generated increases daily, which makes it difficult to process. In the case of supervised learning, labeling training examples may represent an especially tedious and costly task. One of the aims of positive and unlabeled (PU) learning is to train a binary classifier from partially labeled data, representing a strategy for combining supervised and semi-supervised learning and alleviating the cost of labeling data fully. Still, the main strength of PU learning arises when the negative data are not directly available or too diverse. Although the generative approaches have shown promising results in this field, they also bring shortcomings, such as high computational cost, training instability, and inability to generate fully labeled datasets. In the paper, we propose a novel Conditional Generative PU framework (CGenPU) with a built-in auxiliary classifier. We develop a novel loss function to learn the distribution of positive and negative examples, which leads to a unique, desirable equilibrium under a nonparametric assumption. Our CGenPU is evaluated against existing generative approaches using both synthetic and real data. The characteristics of various methods, including ours, are depicted with different toy examples. The results demonstrate the state-of-the-art performance on standard positive and unlabeled learning benchmark datasets. Given only ten labeled CIFAR-10 examples, CGenPU achieves classification accuracy of 84%, while current state-of-the-art D-GAN framework achieves 54%. On top of that, CGenPU is the first single-stage generative framework for PU learning.

Keyword:

Positive and unlabeled learning

Partially supervised learning

Generative adversarial networks

Deep learning