Supporting Database Constraints in Synthetic Data Generation based on Generative Adversarial Networks

With unprecedented development in machine learning algorithms, it is crucial to have available large amount of data to verify the correctness and efficiency of these algorithms. Due to privacy concerns, we may not always have enough real data to use. In our research, we focus on data synthesization for relational databases where the database constraints of the original data must be imposed to the generated data. To the best of our knowledge, no study has been conducted on supporting database constraints in synthetic data generation. We offer solutions by designing extensions to Tabular Generative Adversarial Network algorithm. We implemented a prototype for our approach, and compared the performance of different extensions by experiments. Related work on synthetic data generation includes classical statistical methods and neural network approaches. Synthetic Data Vault is developed using classical statistical methods. It uses Kolmogorov-Smirnov test to select the best statistical distribution to describe columnar data. TableGAN and Tabular GAN use neural networks to minimize cross entropy or Kullback-Leibler divergence on marginal distributions. The main challenges to our research problem are: Classical statistical distributions cannot describe complex and mixed distributions in relational databases. Database constraints are non-differentiable. Neural networks require loss functions to be differentiable.