One-dimensional data augmentation using a wasserstein generative adversarial network with supervised signal

Abstract

In recent years, Generative Adversarial Network (GAN) is widely applied in many domains; however, there is still some difficulties in training the network, which are mainly caused by mode collapse, vanishing gradient of generator and indirect assessment criteria of generated samples. In this paper, the supervised signal is introduced into Wasserstein Generative Adversarial Network (WGAN) on the application of one-dimensional data augmentation to alleviate this difficulty. In the proposed method, besides generating fake samples, a well trained generative model is implemented to reconstruct the real samples , whose input data are latent space samples obtained from autoencoder (AE). In addition, the mode collapse can be prevented by the new model through ensuring that the supervised signal grounded in all the available training data. The performance of our method is verified based on parameters of electronic equipment and stock index systematically and quantitatively, and the superiority of the algorithm is demonstrated by the experiment results both in convergence rate and the quality of samples compared with WGAN and VAEGAN.