Fair synthetic tabular data generation via disentangled semantic encoding and large language models

Abstract Synthetic tabular data generation plays a key role in alleviating data scarcity and privacy limitations in machine learning, especially in regulated domains such as finance and healthcare. While many existing methods focus on producing realistic and predictive tabular data, fairness is often not explicitly considered, leading to biased synthetic datasets. Recent approaches based on large language models (LLMs) further enhance data quality through natural language representations, which underscores the need for principled mechanisms that ensure fairness alongside predictive performance. To address this challenge, we propose SENNA, a framework that integrates disentangled semantic representations with LLM-based generation to jointly preserve utility and improve fairness. The framework separates task-relevant and task-independent information in the latent space and leverages these representations to guide generation and validation. Experimental results on three real-world datasets show that SENNA achieves competitive downstream performance, while substantially reducing both group-level disparities and individual-level inequality. We further evaluate the framework across multiple backbones, providing insights into fairness–utility trade-offs in LLM-based tabular data synthesis.