Distribution Enhancement for Imbalanced Data with Generative Adversarial Network

Tackling imbalanced problems encountered in real-world applications poses a challenge at present. Oversampling is a widely useful method for imbalanced tabular data. However, most traditional oversampling methods generate samples by interpolation of minority (positive) class, failing to entirely capture the probability density distribution of the original data. In this paper, a novel oversampling method is presented based on generative adversarial network (GAN) with the originality of introducing three strategies to enhance the distribution of the positive class, called GAN-E. The first strategy is to inject prior knowledge of positive class into the latent space of GAN, improving sample emulation. The second strategy is to inject random noise containing this prior knowledge into both original and generated positive samples to stretch the learning space of the discriminator of GAN. The third one is to use multiple GANs to learn comprehensive probability distributions of positive class based on multi-scale data to eliminate the influence of GAN on generating aggregate samples. The experimental results and statistical tests obtained on 18 commonly used imbalanced datasets show that the proposed method comes with a better performance in terms of G-mean, F-measure, AUC and accuracy than 14 other rebalanced methods. This paper introduces three strategies to improve the ability of GAN to handle imbalanced data. The first strategy is to inject prior knowledge into the latent space of GAN. The second strategy is to inject random noise into the discriminator. The third one is to use multiple GANs to learn comprehensive probability distributions of positive class based on multi-scale data. image