Adversarial Training Classifier for Imbalanced and Semi-Supervised Learning

The imbalanced learning research has made great progress due to the introduction of generative adversarial networks (GANs). However, most studies focus on combining GAN models with oversampling techniques which could potentially compromise the distribution of the original training dataset. This study presents a cost-sensitive classifier based on the adversarial training framework that can not only deal with imbalanced data distribution but also utilize the unlabeled sample sets for semi-supervised learning. Extensive experiments are carried out to compare the classification performance on imbalanced datasets with the resampling and balanced ensemble method, as well as class-imbalanced semi-supervised scenarios, to demonstrate the advantages of the proposed method.

I. Introduction

THE class imbalance problem exists in almost every real-world scenario, which refers to the phenomenon that different classes are not equally represented in a dataset due to large gaps in sample size. The collected data is inherently imbalanced in real-world applications, such as fault diagnosis [1]–[3], medical diagnosis [4]–[6], anomaly detection [7], [8], and object detection [9], [10], among others. The majority classes would undermine the decision boundary, causing the misclassification of minority samples [11]. But minority samples have greater value for the application task, if not detected, enormous damage could be done because of a higher misclassification cost compared to majority ones. Moreover, because it's very expensive to manually label all of the training data, recent studies also propose countermeasures for class-imbalanced semi-supervised learning scenarios [12]–[14].