A new PCA-based utility measure for synthetic data evaluation

Abstract

Data synthesis is a privacy enhancing technology aiming to produce realistic and timely data when real data is hard to obtain. Utility of synthetic data generators (SDGs) has been investigated through different utility metrics. These metrics have been found to generate conflicting conclusions making direct comparison of SDGs surprisingly difficult. Moreover, prior research found no correlation between popular metrics, concluding they tackle different utility-dimensions. This paper aggregates four popular utility metrics (representing different utility dimensions) into one using principal-component-analysis and checks whether the new measure can generate synthetic data that perform well in real-life. The new measure is used to compare four well-recognized SDGs.