| Abstract: |
Synthetic tabular data generation is increasingly used in
healthcare to augment limited datasets, enable data sharing, and support model development. Current state-of-the-art methods for tabular
synthetic data generation are dominated by deep generative models, including GAN- and VAE-based approaches such as CTGAN and TVAE,
and more recent diffusion-based models such as TabDDPM. These methods can reproduce complex data distributions with high fidelity, but they
do so through black-box mechanisms: a trained generator consists of
learned network parameters rather than an explicit, traceable procedure
for producing each synthetic feature value.
In a clinical setting, where data auditability and clinician trust are central concerns and regulatory frameworks increasingly call for explainable
AI, this opacity is a meaningful limitation, independent of whatever fidelity scores the model achieves. In this paper, we propose a generator built from Cartesian Genetic Programming (CGP): a population of
symbolic computational graphs evolved directly against a distributionalfidelity objective. Each evolved generator is a finite, explicit graph of
mathematical operations that can be traced and audited feature by feature. We describe the genome representation, the (1 + λ) evolutionary
search procedure, a mixed numerical/categorical fitness function based
on Kolmogorov–Smirnov and total variation distance, and a multi-axis
evaluation protocol covering fidelity, downstream classification utility,
and privacy. We evaluate the proposed generator across three tabular
datasets covering heart disease, fetal health, and wine quality, and analyze the tradeoff between symbolic interpretability and the performance
of black-box deep generative baselines. |