One software defective data augmentation method based on VAE and WGAN

In software metric datasets, the number of defective samples is always even fewer than that of non-defective samples, which makes follow-up research complex and difficult. Therefore, this essay provides a method of software defective data augmentation based on a variational autoencoder (VAE) and Wasserstein Generative Adversarial Network (WGAN). This process contains several steps. First, dimensionality reduction of software metric data is achieved by utilizing VAE, producing a set of codes (latent vectors); the distribution of the set is studied by WGAN and then “the code set (latent vectors)” is generated. Finally, the generated codes (latent vectors) are input into VAE to acquire defective data. This paper proposes a data augmentation method of a minority class, based on VAE and WGAN. The experiments performed in MINST, NSAS MDP confirm that 1) This data augmentation method is superior to WGAN, AE+WGAN, and SMOTE; 2) Introducing the variance of generated samples to the WGAN generator, the variety of those is efficiently improved.