Learning with Missing Data

The field studies how to use incomplete observations without introducing misleading estimates or discarding useful examples, especially when the chance of missingness depends on the data itself. Approaches range from likelihood-based methods, EM, and multiple imputation to Bayesian and diffusion generative models, masked transformers, and predictors that encode missingness patterns directly. Work spans tabular data, irregular time series, clinical records, graphs, and multimodal settings, with theory addressing identifiability, sample complexity, and uncertainty under MCAR, MAR, and MNAR mechanisms. A central tension is that missingness can be both a useful signal and a source of bias: methods must match the missingness process and downstream task, while evaluations increasingly test realistic patterns and train–test shifts rather than reconstruction error alone.