Learning from Class-Imbalanced Data Using Misclassification-Focusing Generative Adversarial Networks

• DOI: 10.2139/ssrn.4342079
• Publication Date: 2023-01-31T20:21:10Z
• Authors: Jaesub Yun; Jong-Seok Lee

This paper presents a novel end-to-end oversampling-classification approach, which we refer to as imbalanced data-classifying generative adversarial network (ImbGAN), for data classification. ImbGAN has classifier-embedded structure within GAN and consists of five components: (1) generator, (2) discriminator, (3) classifier, (4) storage misclassified minority class data, (5) artificial data. By iterative interaction with the embedded first two components generate instances that are similar by classifier. Therefore, these three networks iteratively simultaneously trained. The stored in fourth fifth components, respectively. These also updated iterations proceed. Our method obtains final classification model from single learning process, while most generation methods go through an additional process training classifiers after generation. Numerical experiments based on tabular, image, text datasets confirm proposed outperforms well-known synthetic sampling methods, such SMOTE, GAN, their variants.