Data 001: Synthetic Data Under Data Scarcity
A benchmark testing whether Gaussian Copula synthetic data can make up for scarce real healthcare training data. It couldn’t: augmentation helped 0 of 18 conditions and hurt 14.
Data Scientist, Synthetic Data Under Data Scarcity, 2026.
The idea
When real data is scarce, can synthetic data fill the gap? This project tests whether Gaussian Copula augmentation can make up for missing real healthcare training data.
What I did
- Compared Gaussian Copula augmentation against real-only training on two healthcare datasets (Diabetes and Cleveland Heart Disease) with Logistic Regression and Random Forest
- Simulated scarcity from 50% down to 1% of the real training data
- Ran 30 paired, repeated splits per condition, always testing on untouched real data
- Predefined a ±0.01 ROC-AUC practical-effect threshold and ran fidelity diagnostics
- Froze and verified the release with a reproducibility script and pytest
Result
Augmentation helped 0 of 18 conditions and hurt 14, so synthetic rows did not replace missing real data.