Orthogonal Procrustes problem preserves correlations in synthetic data
Researchers have developed a new postprocessing technique for synthetic tabular data that uses the Orthogonal Procrustes problem to restore the original data's Pearson correlation structure. This method aims to preserve the dependence structure, which is crucial for applications involving privacy, data sharing, and scarcity. Experiments show that the approach effectively restores correlations while maintaining individual feature distributions, data geometry, and downstream classification task performance. AI
IMPACT Enhances the utility of synthetic data by preserving its statistical properties, potentially improving privacy-preserving AI development.