Missingness-aware Data Imputation via AI-powered Bayesian Generative Modeling
New model quantifies uncertainty over missing values, not just point estimates.
Missing data plagues many real-world datasets, but most imputation methods only give a single best guess. Qiao Liu's new work, MissBGM, tackles this by combining neural network flexibility with Bayesian inference rigor. The model jointly learns the data-generating process and the mechanism responsible for missing values, producing full posterior distributions over imputations rather than point estimates. A stochastic optimization framework cycles through updates to missing values, model parameters, and latent variables until convergence, with theoretical guarantees that estimates converge under mild assumptions.
In empirical tests across diverse settings, MissBGM consistently beat traditional techniques like MICE and more recent neural imputers. The method's ability to output uncertainty intervals is critical for high-stakes applications in healthcare, finance, and scientific research, where knowing what you don't know is just as important as the imputed value itself. Liu has open-sourced the code, enabling other researchers and practitioners to adopt or build upon this approach. The paper is currently on arXiv and has drawn attention for bridging a key gap between statistical rigor and deep learning scalability.
- MissBGM explicitly models both the data-generating and missingness mechanisms, unlike point-estimate methods.
- Stochastic optimization alternates updates across missing values, parameters, and latent variables with theoretical convergence guarantees.
- Outperforms traditional imputers and recent neural network methods across extensive experimental settings; code is open source.
Why It Matters
Gives data scientists principled uncertainty for imputations, critical for high-stakes decisions in healthcare, finance, and research.