A Generative Model for Synthesizing Structured Datasets Based on GAN

Synthesizing high quality dataset has been a long-standing challenge in both machine learning and database community. One of the applications of high quality dataset synthesis is to improve the model training, especially deep learning models. A robust model training process requires a large annotated dataset. One way of acquiring a large annotated training set is via the domain experts manual annotation, which is expensive and prone to mistakes. Therefore, as an alternative, automatic synthesis of high quality and similar dataset is much more plausible. Some efforts have been devoted for synthesizing image dataset due to the rapid development of computer vision. However, those models can not be applied to the structured data (numeric & categorical table) directly. Moreover, little efforts have been payed to the numeric & categorical table. Therefore, we propose TableGAN, the first generative model from GAN family, which improves the performance of the generative model with adversarial learning mechanism. TableGAN modifies the internal structure of traditional GAN targeting numeric & categorical table, including the optimization function, to synthesize more high-quality training dataset samples for improving the effectiveness of the training models. Extensive experiments on real datasets show significant performance improvement for those models trained on the enlarged training datasets, and thus verify the effectiveness of our TableGAN.