ETGAN: A Hybrid GAN Ensemble for Synthesizing Time-Dependent and Static Tabular Data

The increasing reliance on data-driven decisions across industries has accentuated the importance of machine learning models. However, these models often encounter challenges that can be effectively addressed by ensuring the provision of quality data in large quantities. This is where GANs are thriving as they help to tackle data shortage and privacy issues by generating synthetic data that is statistically similar to the original data. Enterprise data is mostly tabular and the existing GAN algorithms have limited focus on tabular data. Most of these GANs focus on only static or point-in-time tabular data and very few of them focus on time series or dynamic data. Therefore, there is a crucial need for a generic model capable of generating both static and time-dependent features in tabular data, given that real-world business problems involve both static and time-dependent or time-series values. In this paper, by using GAN architecture embedded with a transformer, we propose a novel methodology to generate synthetic data that is a mix of static and time-dependent or time-series features. This work presents underlying mathematical principles and empirical evaluations that demonstrate superior performance compared to existing models, showcasing improved sample quality and diversity. We support our claim with experimental results and conclude that the proposed state-of-the-art methodology is better than the existing architectures.