Dependency-Aware Fidelity Diagnostic for Synthetic Tabular Data

A new paper on arXiv (2607.21636) reveals that common metrics for synthetic tabular data, such as logistic-regression C2ST and pairwise Trend score, are largely blind to inter-column dependencies. The authors introduce a dependency-aware diagnostic that decomposes an XGB-C2ST test into marginal, dependency, and cross components, anchored between a fully-factorized reference and a real-data oracle. This method aims to better preserve the discriminative signal for minority classes in imbalanced domains like fraud and clinical risk.