A new research paper explores latent-space flow models for generating synthetic tabular data, focusing on balancing analytical utility with disclosure risk and computational cost. The study investigates various learning targets, probability paths, and sampling methods across seven datasets. Key findings indicate that learning targets significantly influence the utility-risk trade-off, with velocity and posterior matching favoring higher utility and score/noise matching prioritizing lower risk. Configuration choices, such as path selection and integration budgets, also impact performance, offering potential for compute savings. AI
IMPACT Provides guidance for selecting pre-release models for synthetic tabular data generation under resource constraints.
RANK_REASON Academic paper published on arXiv detailing a new methodology for synthetic data generation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
