Machine Learning
ReFine: A Two-Part Framework for Generating Synthetic Tabular Data from Limited Records
Researchers propose a new approach that combines symbolic rules and filtering to generate synthetic tabular data reliably when real-world datasets are scarce.
A new research paper from arXiv presents ReFine, a framework designed to tackle a persistent challenge in machine learning: generating high-quality synthetic tabular data when only limited reference data is available.
The Core Problem#
Existing approaches like generative adversarial networks (GANs) and fine-tuned large language models typically require substantial real-world data to work effectively. Prompt-based LLMs offer more flexibility without parameter tuning, but they often produce synthetically generated data that drifts from the true distribution and contains localized redundancy—patterns that harm downstream task performance.
The ReFine Solution#
The framework combines two complementary components:
-
Symbolic Rule Extraction & Guidance: ReFine extracts if-then rules from interpretable models and embeds them directly into generation prompts. This steers the LLM toward domain-specific distributions rather than generic patterns.
-
Dual-Granularity Filtering: A filtering mechanism simultaneously reduces over-sampling of common patterns while preserving rare but valuable samples, eliminating the localized redundancy problem.
Results#
Experiments across diverse benchmarks show ReFine achieves top-tier performance consistency. In extreme low-data regimes, the framework delivers an average relative improvement of 7.48% in downstream utility across tested datasets.
The work addresses a real gap in synthetic data generation: most existing methods assume abundant reference data, which isn't always realistic for domain-specific applications or proprietary datasets.
Source: arXiv cs.AI