RESEARCH · RESEARCH · #3697
Understanding Private Evolution as Learning‑Augmented Clustering (NeurIPS 2024)
Authors Audra McMillan, Kunal Talwar and Felix Zhou recast Private Evolution (PE) — a differentially private algorithm for synthetic data — as generative-model‑augmented Wasserstein learning. They prove improved sample-complexity bounds when a generator captures low‑dimensional structure, show standard PE variants can fail to converge on simple clustered instances, and propose a new geometry‑aware PE with provable convergence; experiments report competitive performance and improved recall.
KEY POINTS
- Authors Audra McMillan, Kunal Talwar and Felix Zhou recast Private Evolution (PE) — a differentially private algorithm for synthetic data — as generative-model‑augmented Wasserstein learning.
- They prove improved sample-complexity bounds when a generator captures low‑dimensional structure, show standard PE variants can fail to converge on simple clustered instances, and propose a new geometry‑aware PE with provable convergence; experiments report competitive performance and improved recall.
- Recasting PE to account for generators and adding a geometry‑aware variant gives stronger theoretical guarantees and practical fixes for convergence and sample efficiency in private synthetic-data generation.
WHY IT MATTERS
Recasting PE to account for generators and adding a geometry‑aware variant gives stronger theoretical guarantees and practical fixes for convergence and sample efficiency in private synthetic-data generation.