Time Series Imputation

Time series imputation estimates unobserved values in sequences such as medical measurements, sensor readings, weather, and energy use, so they can be analyzed or used for forecasting and decisions. Approaches range from interpolation and low-rank statistical models to recurrent networks, Transformers, graph models, and probabilistic methods such as diffusion and flow matching; continuous-time models address irregular sampling. A central challenge is that missingness may occur in long blocks or depend on the unobserved values themselves, so methods increasingly model missingness patterns and report uncertainty. Evaluation remains sensitive to the gap pattern and data regime: low average error can hide damage to important signal structure, and simple methods may work well in stable intervals while learned models help more during complex transients.