AugMask Fixes AI Data Generators That Choke on Real-World Missing Data
AugMask lets any diffusion-based tabular generator train on incomplete data without architectural changes, beating specialized missing-aware models
Takeaways
- The problem: Diffusion-based tabular generators assume fully observed data, but real datasets almost always have missing values.
- The solution: AugMask is a plug-and-play training wrapper that fills missing cells with a stochastic auxiliary model and masks the loss to only observed coordinates.
- Key insight: Missing entries are used as uncertain conditioning context, not as training targets, preventing the model from learning fabricated signal.
- Results: AugMask-wrapped CDTD matches or beats specialized missing-aware baselines (ForestDiff, MIWAE, MissDiff) across 90 benchmark scenarios with missing ratios from 10% to 90%.
- Accepted at ICML 2026 (PMLR 306); code release is forthcoming.
- Limitation: AugMask is a generation framework only -- it does not directly impute missing values in the original dataset, and high-dimensional and privacy settings are left for future work.
Most real-world tabular datasets are riddled with holes. A hospital record missing lab values, a financial dataset with unreported fields, a survey with skipped questions -- incomplete data is the norm, not the exception. Yet the best tabular data generators, which are built on score-based diffusion models, were designed assuming every row is fully filled in. AugMask, accepted at ICML 2026, is a plug-and-play training wrapper that fixes this mismatch without touching a model's architecture.
The wall every tabular generator hits
Deep generative models have been successfully applied to synthesize tabular data, addressing challenges like data scarcity and privacy, but real-world tabular data often contain missing values, hindering the direct application of these models to practical settings. The standard workaround is the "impute-then-generate" pipeline: fill in the blanks with a separate imputation model first, then train the generator on the completed data. The widely adopted "impute-then-generate" pipeline may lead to a biased learning objective -- the generator learns from a version of the data that was partly invented, and that invented signal leaks into the learned distribution.
The vanilla diffusion model requires complete or fully observed data for training, and incomplete data is a common issue in various real-world applications, including healthcare and finance, particularly when dealing with tabular datasets. Prior specialized approaches like MissDiff and MIWAE tried to solve this by redesigning the model itself -- masking the loss directly, or building variational objectives that marginalize over missing values. These work, but they require you to swap out your generator entirely.
Two moves, one wrapper
AugMask is a plug-and-play training framework that adapts missing-unaware backbones to incomplete data by separating conditioning from supervision. The key insight is elegant: instead of forcing the model to ignore missing values or pretending they don't exist, AugMask treats them as uncertain context -- useful signal for shaping the forward pass, but never something the model is graded on.