Incomplete Data Imputation

Incomplete data imputation studies how to recover values that were not recorded, so datasets can support analysis, prediction, and decision-making without treating guesses as facts. Approaches range from multiple imputation by chained equations and low-rank matrix completion to graph and time-series models, variational autoencoders, diffusion models, and flow matching. A central challenge is that missingness may depend on the unseen value itself (MNAR), so methods increasingly model the missingness process and report uncertainty, not just a single estimate. No method works best everywhere, and low reconstruction error does not always translate into reliable downstream conclusions, especially when missingness patterns change between training and deployment.