Title: AugMask: Adapting Diffusion Models to Incomplete Tabular Data Through Stochastic Augmentation and Masking

Abstract:

While score-based diffusion models have established themselves as leading deep generative architectures, their utility for tabular data is hindered by a fundamental mismatch: their underlying structures typically require fully specified inputs, yet practical tabular datasets frequently suffer from missing values. To address this, we introduce AugMask, a modular training framework designed to retrofit missing-unaware backbones for use with incomplete data by decoupling the conditioning process from supervision.

AugMask operates through two primary mechanisms. First, it generates numeric inputs by employing lightweight auxiliary models to perform conditional stochastic augmentation. Second, it restricts denoising supervision exclusively to the observed data points. Consequently, the augmented values for missing entries function as a context of uncertainty for conditioning, rather than serving as direct targets for training. We demonstrate that this training protocol aligns with a Rao--Blackwellized objective, where marginalizing over missing entries introduces a penalty based on variance-weighted sensitivity. This mechanism effectively prevents the model from becoming overly dependent on uncertain imputations. Empirical results across various datasets and missingness scenarios indicate that AugMask allows standard diffusion-based generators for tabular data to surpass specialized baselines designed to handle missing information.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC