Reproduce “ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaofeng Lin, Seungbae Kim, Zhuoya Li, Zachary DeSoto, Charles Fleming, Guang Cheng
ReTabSyn, a framework developed at the University of California, Los Angeles and the University of South Florida, synthesizes realistic tabular data by aligning deep generative models with feature-target conditional distributions using Direct Preference Optimization and oracle-free preference labeling. It consistently improves downstream machine learning utility in low-data, imbalanced, and distribution-shifted scenarios while maintaining strong statistical fidelity and reduced privacy leakage compared to interpolation methods.
Key insights
- Maximizing downstream model utility for tabular data is primarily driven by accurately modeling the conditional distribution P(y|X), rather than perfectly replicating the full joint distribution P(X,y).
- High-purity preference signals for conditional alignment can be automatically generated using rule-based perturbations derived directly from tabular schemas, eliminating the need for external oracles or human labels.
- Direct Preference Optimization (DPO) can effectively fine-tune tabular generative models to prioritize decision-relevant structures, leading to synthetic data with improved utility, particularly in challenging data regimes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.