Improved Density Ratio Estimation for Evaluating Synthetic Data Quality

High-quality synthetic data is essential for accurate downstream analysis. Density Ratio Estimation (DRE) has emerged as a powerful tool for evaluating synthetic data quality. However, existing DRE methods are highly sensitive to hyperparameter selection, where suboptimal choices lead to poor convergence rates and degraded empirical performance. To mitigate this, we propose a novel model aggregation algorithm for DRE that trains multiple models with diverse hyperparameter configurations and combines their outputs. Our approach achieves fast convergence without requiring prior knowledge of the unknown density ratio smoothness and is minimax optimal for the squared loss. We demonstrate that our method enhances the performance of established DRE techniques across benchmark datasets, achieving state-of-the-art results on MiniDomainNet and Amazon Reviews.