Enhancing Tabular Data Generation With Dual-Scale Noise Modeling

The generation of synthetic tabular data plays a critical role in applications such as data enhancement, privacy preservation, and model validation. However, the heterogeneity of tabular data—comprising both continuous and categorical features with intricate interdependencies—presents significant challenges in generating high-quality data. This study presents the Dual-Scale Diffusion Probabilistic Method (DSDDPM), a novel framework developed to improve the Denoising Diffusion Probabilistic Model (DDPM) performance for tabular data generation. DSDDPM tackles the challenge of capturing both global dependencies and local details through the use of a dual-scale noise-handling mechanism. It models the global structure at the coarse scale and the regional patterns at the fine scale throughout the forward and backward diffusion processes. Experimental evaluations on several benchmark tabular datasets show that DSDDPM outperforms existing state-of-the-art methods. Specifically, DSDDPM improves machine learning efficiency by an average of 5-7% compared to other methods. DSDDPM exhibits the lowest correlation deviation across all datasets regarding global dependency modeling. Moreover, regarding detail consistency, DSDDPM shows the highest alignment of feature distributions between the generated and actual data, significantly outperforming other methods. These results position DSDDPM as a robust and scalable solution for generating high-quality synthetic tabular data, enhancing fidelity and versatility across various applications.