Generating High-Quality Tabular Synthetic Data at Scale
summary
Introduces GEM+ as a generator‑network replacement for probabilistic graphical models that scales synthetic tabular generation to hundreds and up to ~1,000 columns, claiming a 10% improvement in accuracy over AIM on realistic privacy budgets (epsilon ≥ 0.5) and lower compute. Describes a vertical public/private split workflow: pretrain on public marginals with no DP noise, allocate privacy budget only to private columns, and use conditional generation on public fields to boost quality. Reports retaining AB‑test outcomes ~95% of the time and competitive ML performance for feature expansion and full replacement.
AI-written summary · about
chapters (8)
- 0:00Synthetic data generation intro
- 1:58Why synthetic data needed
- 4:05AIM algorithm limits
- 6:10Scaling with generator networks
- 8:15Gen Plus vs AIM
- 11:06Public and private data splits
- 13:37Synthetic data applications
- 15:25Summary & Q&A
related talks
45mPractical Experience with Making Synthetic Data Differentially Private
16mCompute Engine Testing with Synthetic Data Generation
25mThe Importance of Synthetic Data in Data-Centric AI
01h 14mStatistical Inference and Privacy, Part III
36mGenerating realistic, differentially private data sets using GANs
01h 00mThe Synthetic Data Advantage: How It Helps You to Get Both AI Privacy & Explainability
30mPrivate Generative Models: Balancing Privacy and Utility in AI
30m🤯 No data? No problem! Synthetic data to the rescue