TabMAGD: Combining Masked Autoencoders with Guided Diffusion for Mixed-Type Tabular Data Synthesis

To address the scarcity of real data and data privacy concerns, significant research efforts have been directed toward developing tabular data synthesizers. Tabular data exhibit intrinsic heterogeneity with inter-column correlations, where features may represent diverse information types and contain mixed data formats—including both categorical and numerical values. To overcome challenges in modeling heterogeneous features and their complex dependencies, we propose a novel Tabular data synthesis framework combining Masked Autoencoder with Guided Diffusion (TabMAGD). Unlike existing tabular data generators, our method focuses on the correlation between categorical and numerical features, which learns the impact of numerical features on categorical features through masked autoencoders and captures the influence of categorical features on numerical features through guided diffusion. In our experiments, we conduct a comprehensive evaluation of TabMAGD, demonstrating its state-of-the-art performance against existing generative models. We evaluate TabMAGD’s performance in privacy-sensitive applications and observe that it consistently generates high-quality synthetic data while maintaining an outstanding trade-off between data utility and privacy preservation. Notably, tabular data synthesis can be regarded as a tool for data governance, assisting organizations in generating high-quality, compliant, and secure synthetic data.