Beyond Distributional Fidelity: Causal-Penalized Diffusion for Synthetic Tabular Data
The authors address a flaw in generating synthetic tabular data: optimizing for statistical similarity doesn't guarantee preserving causal effects. They proposed a causal-fidelity-aware training framework that adds a penalty to standard generative objectives. This method successfully improves the preservation of causal effects, such as treatment effects, while keeping the synthetic data statistically similar to real data.