ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning

Deep generative models can help with data scarcity and privacy by producing synthetic training data, but they struggle in low-data, imbalanced tabular settings to fully learn the complex data distribution. We argue that striving for the full joint distribution could be overkill; for greater data efficiency, models should prioritize learning the conditional distribution $P(y\mid \mathbf{X})$, as suggested by recent theoretical analysis. Therefore, we overcome this limitation with ReTabSyn, a Reinforced Tabular Synthesis pipeline that provides direct feedback on feature correlation preservation during synthesizer training. This objective encourages the generator to prioritize the most useful predictive signals when training data is limited, thereby strengthening downstream model utility. We empirically fine-tune a language model-based generator using this approach, and across benchmarks with small sample sizes, class imbalance, and distribution shift, ReTabSyn consistently outperforms state-of-the-art baselines. Moreover, our approach can be readily extended to control various aspects of synthetic tabular data, such as applying expert-specified constraints on generated observations.

Many high-stakes decisions, such as who is at risk of a disease, which customers may leave a service, or which students need extra support, are guided by computer programs that learn from tables of data (like a spreadsheet with one row per person), yet such data is often too sensitive to share and too scarce for the rare cases that matter most. A common fix is "synthetic" data, meaning artificial records that behave like real ones but belong to no real person; the trouble is that most methods copy the data's general look while losing the subtle links between a person's characteristics and the outcome we want to predict, which is exactly what makes the data useful. Our method teaches a data generator to preserve those links by showing it pairs of examples (one realistic and one deliberately altered to break a real-world pattern) and training it to prefer the realistic one, a form of learning from feedback similar to how today's chatbots are refined; crucially, it needs no separate "judge" program, which tends to be unreliable when data is scarce. The resulting synthetic data yields more accurate predictions, especially when real data is small, important cases are rare, or conditions change between where data is collected and where it is used, while revealing less private information than methods that essentially memorize real records and respecting common-sense rules. This is a step toward sharing sensitive information more safely in fields such as healthcare, finance, and the social sciences.