[Paper Review] Synthetic data, real errors: how (not) to publish and use synthetic data

Boris van Breugel, Zhaozhi Qian+1 more|arXiv (Cornell University)|May 16, 2023

Time Series Analysis and Forecasting16 citations

TL;DR

The paper shows that treating synthetic data as real yields poor downstream models and uncertainty, and introduces Deep Generative Ensemble (DGE) to better capture generative uncertainty by training models on multiple synthetic datasets.

ABSTRACT

Generating synthetic data through generative models is gaining interest in the ML community and beyond, promising a future where datasets can be tailored to individual needs. Unfortunately, synthetic data is usually not perfect, resulting in potential errors in downstream tasks. In this work we explore how the generative process affects the downstream ML task. We show that the naive synthetic data approach -- using synthetic data as if it is real -- leads to downstream models and analyses that do not generalize well to real data. As a first step towards better ML in the synthetic data regime, we introduce Deep Generative Ensemble (DGE) -- a framework inspired by Deep Ensembles that aims to implicitly approximate the posterior distribution over the generative process model parameters. DGE improves downstream model training, evaluation, and uncertainty quantification, vastly outperforming the naive approach on average. The largest improvements are achieved for minority classes and low-density regions of the original data, for which the generative uncertainty is largest.

Motivation & Objective

• Demonstrate that naive use of synthetic data leads to poor generalization and unreliable evaluation.
• Introduce Deep Generative Ensemble (DGE) to approximate the posterior over generative model parameters.
• Show that DGE improves downstream model training, evaluation, and uncertainty quantification.
• Highlight how DGE performs particularly well in low-density and minority regions.
• Provide practical guidelines for synthetic data publishers and users.

Proposed method

• Define the downstream task distribution incorporating the generative process via p(T|Dr) and its components.
• Propose Deep Generative Ensemble (DGE) by training K independent generative models and using the empirical distribution over parameters to generate multiple synthetic datasets.
• Use Monte Carlo sampling over (θ, Ds, T) to estimate downstream statistics such as mean and variance.
• Evaluate downstream performance on real data for a range of datasets, comparing naive single-dataset training to DGE.
• Analyze model evaluation, model selection, and uncertainty quantification under synthetic-data regimes.
• Illustrate robustness to overfitting/underfitting of the generator and the impact on underrepresented groups.

Figure 1 : Synthetic data is not perfect, which affects downstream ML tasks, e.g. training a prediction model. The naive synthetic data approach generates one synthetic dataset and treats it like it is real. We propose using an ensemble of generative models for capturing the generative uncertainty,

Experimental results

Research questions

• RQ1 How does treating synthetic data as real affect downstream model performance and uncertainty?
• RQ2 Can a multi-dataset synthetic framework (DGE) better approximate the true posterior over generative model parameters?
• RQ3 Does publishing multiple synthetic datasets improve downstream evaluation, model selection, and uncertainty quantification compared to naive methods?
• RQ4 How does generative uncertainty influence performance in low-density or minority regions?
• RQ5 What practical guidelines should data publishers and users follow when publishing and using synthetic data?

Key findings

• Training on an ensemble of synthetic datasets (DGE) yields real-data-like performance versus naive single-dataset training.
• Naive evaluation overestimates real-world performance, especially as the generator overfits, while DGE provides more conservative and robust estimates.
• DGE better preserves real-world model rankings for downstream tasks and reduces selection bias toward overly complex models.
• DGE improves uncertainty quantification by capturing generative uncertainty, aligning predictive uncertainty with generative variability.
• Performance gains from DGE are largest in low-density/minority regions and when the generator is imperfect.
• Publishing separate synthetic datasets (with metadata) enables proper estimation of generative uncertainty.

Figure 2 : Conclusions drawn from synthetic data do not always transfer to real data.