Data 002: Synthetic Restoration Controls under Data Scarcity
A matched-control follow-up to Data 001 that tests whether Gaussian Copula augmentation beats simply duplicating scarce real rows. It never did, and it was practically harmful in 10 of 18 settings.
Data Scientist, Synthetic Restoration Controls under Data Scarcity, 2026.
The idea
Data 001 showed synthetic rows hurt under scarcity, but maybe only because the table was padded back to full size. This follow-up compares Gaussian augmentation against a simple control: duplicating the real rows.
What I did
- Designed a 540-condition matched-control study holding training size, class balance, splits, models, and test sets constant
- Compared three training setups: scarce real-only, replication-restored, and Gaussian Copula-augmented
- Estimated uncertainty with 10,000 paired bootstrap resamples over 30 seeds per setting
- Built a reproducible, hash-verified pipeline with a 98-test validation suite
Result
Gaussian augmentation was never better than replication: harmful in 10 of 18 settings, with no clear difference in the other 8.