Synthetic Tabular Data Generation

The field studies how to create artificial tables with realistic numerical and categorical values, dependencies, and—in some settings—relational or temporal structure. Approaches include GANs and VAEs, copulas and Bayesian networks, diffusion and flow models, tree-based generators, and language-model methods, often augmented with causal knowledge or hard domain constraints. Synthetic tables can support data sharing, model training, rare-event augmentation, and analysis when real records are scarce or sensitive; evaluation therefore considers statistical fidelity, downstream utility, privacy, and fairness. A central tension is that matching marginals or predictive performance may still miss dependencies, rare patterns, causal effects, and valid domain rules, while complex deep generators do not consistently outperform simpler statistical baselines, especially on small datasets.