RDDMPI: Residual Denoising Diffusion for Probabilistic Multivariate Time Series Imputation

Probabilistic multivariate time series imputation often overwhelms diffusion models by forcing them to model global dynamics and fine noise simultaneously. RDDMPI addresses this by shifting diffusion directly into residual space. A deterministic pre-trained imputer first reconstructs the baseline trajectory, after which a conditional residual diffusion process refines uncertainty, modulated by reliability-aware conditioning. Benchmark evaluations show marked gains in both deterministic accuracy and probabilistic calibration.