Tabular Diffusion Transformer for Cross-Table Synthetic Data Generation

A new research paper on arXiv (2608.14496) introduces a two-stage framework for cross-tabular data generation (CTDG), addressing the limitations of existing methods that are restricted to single-input-table scenarios. The first stage transforms each heterogeneous raw table into a standardized statistical table with uniform columns, capturing marginal distributions and pairwise correlations. The second stage employs a diffusion transformer model to learn structural patterns across these homogeneous statistical tables and generate synthetic statistical tables, which are then used to produce synthetic raw tables. This approach enables the generation of synthetic data from multiple heterogeneous tables with diverse feature sets, a task previously challenging for existing methods. The paper is authored by researchers and posted on arXiv, a preprint server. The work is significant for the field of synthetic data generation, particularly in healthcare and other domains where data privacy and heterogeneity are concerns. The framework's ability to handle multiple tables with different features could improve the utility and realism of synthetic datasets, facilitating research and development without compromising sensitive information. The paper is available at the provided URL.