A Multi-Dimensional Evaluation of Synthetic Data Generators

Synthetic datasets are gradually emerging as solutions for data sharing. Multiple synthetic generators have been introduced in the last decade fueled by advancement machine learning and increased demand fast inclusive sharing, yet their utility is not well understood. Prior research tried to compare of using different evaluation metrics. These metrics found generate conflicting conclusions making direct comparison very difficult. This paper identifies four criteria (or dimensions) masked classifying available into categories based on measure they attempt preserve: attribute fidelity, bivariate population application fidelity. A representative metric from each category chosen popularity consistency, used overall recent synthesizers across 19 sizes feature counts. The also examines correlations between selected an streamline utility.