Evaluation of Synthetic Data Generators on Complex Tabular Data

This paper evaluates a range of synthetic data generators on complex tabular data — the mixed-type, structured data common in real applications. It compares methods on both syntactic accuracy, whether the synthetic records respect the data’s formats and constraints, and statistical accuracy, how well they reproduce distributions and relationships. By testing several approaches on the same demanding dataset, it clarifies their relative strengths and weaknesses for practical use. Presented at Privacy in Statistical Databases, the study contributes to the SwissAnon group’s systematic benchmarking of synthetic-data methods.