A Comparison Study of VAE and GAN for Software Fault Prediction

Software fault is an unavoidable problem in software project. How to predict software fault to enhance safety and reliability of system is worth studying. In recent years, deep learning has been widely used in the fields of image, text and voice. However it is seldom applied in the field of software fault prediction. Considering the ability of deep learning, we select the deep learning techniques of VAE and GAN for software fault prediction and compare the performance of them. There is one salient feature of software fault data. The proportion of non-fault data is well above the proportion of fault data. Because of the imbalanced data, it is difficult to get high accuracy to predict software fault. As we known, VAE and GAN are able to generate synthetic samples that obey the distribution of real data. We try to take advantage of their power to generate new fault samples in order to improve the accuracy of software fault prediction. The architectures of VAE and GAN are designed to fit for the high dimensional software fault data. New software fault samples are generated to balance the software fault datasets in order to get better performance for software fault prediction. The models of VAE and GAN are trained on GPU TITAN X. SMOTE is also adopted in order to compare the performance with VAE and GAN. The results in the experiment show that VAE and GAN are useful techniques for software fault prediction and VAE has better performance than GAN on this issue.