A Comparative Analysis of GAN and VAE based Synthetic Data Generators for High Dimensional, Imbalanced Tabular data

Synthetic data has emerged as an acceptable solution in machine learning that overcomes the constraints of data availability due to data privacy restrictions. Other major challenge with machine learning is dealing with imbalanced dataset. Several techniques exist to deal with the data imbalance, however, the problem continues to exist when using synthetic data generators dealing with highly imbalanced datasets. Generative Adversarial Network is already proven to be an excellent model to generate synthetic data, especially for high-dimensional datasets. There are other deep learning models that use Variational Autoencoders and Recurrent Neural Networks which are also being explored. To understand how these generators perform when presented situations dealing with highly imbalanced datasets, we experimentally evaluate two deep-learning synthetic data generators, one is based on Generative Adversarial Network (CTGAN) and the other is on Variational Auto Encoder (TVAE). We assess how each of these performs when presented with two datasets of distinct characteristics. The datasets used are high dimensional, highly imbalanced tabular data with one dataset having 19.3% minority class and the other having only 5.68% of minority class. Our test results find that TVAE fails to generate minority data when the minority class is very small in number.