Causal-TGAN: Causally-Aware Synthetic Tabular Data Generative Adversarial Network
Synthetic tabular data generation has recently gained immense attention due to applications in medicine, finance, and other fields. Generative adversarial networks (GANs) designed initially for image generation have been demonstrated to be promising for generating certain types of tabular data. Tabular data may contain mixed data types such as continuous, ordered, binary, and categorical values. However, the causal relationships between the variables in tabular data have been largely ignored by the prior art. Causality encodes real-world relationships occurring naturally between variables measuring a phenomenon. In this work, we propose Causal-TGAN, a data generation architecture that incorporates causal relationships at its core. The flexibility of this architecture is its capability to support different types of expert knowledge (e.g., complete or partial) about the causal nature of the underlying phenomenon. Extensive experimental results on both simulated and real-world datasets demonstrate that Causal-TGAN and its hybrid avatars consistently outperform other baseline GAN models. We also argue that the architecture's flexibility is promising for many practical applications.
PDF Abstract
