Adaptive Constrained Optimization for Tabular Synthetic Data Generation

Synthetic data generation has recently emerged as a solution in data-scarce regulated industries, such as finance and healthcare. While synthetic data requires navigating various tradeoffs implicitly or explicitly, including fidelity, utility, fairness or privacy properties, business objectives are usually focused on a single dimension. Although recent optimization approaches such as SCGOAT (Hamad et al., 2023) enable Bayesian optimization to explore tradeoffs in synthetic data generation, determining the boundaries for constraints remains challenging, as it relies on both the original dataset and the trained generator(s). To tackle this issue, we propose a novel Adaptive Constrained Threshold (ACT) strategy within the SCGOAT framework. Our method starts by relaxing the constraints to identify feasible regions and progressively tightens them, thereby minimizing wasted evaluations in infeasible spaces. Experiments on tabular datasets demonstrate that our approach achieves a competitive tradeoffs between various synthetic data dimensions such as downstream performance, fidelity and privacy, improving over fixed-constraints baseline approaches.