Virtual Sample Generation Based on Regression Enhanced Generative Adversarial Network

The limited information and knowledge inherent in small sample sizes often result in poor generalization performance of constructed models. To address this issue, we propose a novel method for virtual sample generation (VSG) utilizing a regression enhanced generative adversarial network (REGAN). Initially, the variational autoencoder (VAE) serves as the generator within REGAN, with an additional regression layer incorporated to augment the regression information pertaining to virtual samples. Subsequently, both real and virtual samples undergo discrimination and prediction by the discriminator, thereby reinforcing the regression relationship among variables. Ultimately, optimal virtual samples are selected utilizing a co-training strategy. The efficacy of our proposed method is validated through experimentation on benchmark datasets.