RDDMPI: Residual Denoising Diffusion Model for Probabilistic Multivariate Time Series Imputation

Recent diffusion-based approaches to multivariate time series imputation (MTSI) model the full data distribution via iterative denoising, requiring the network to simultaneously capture global structure, temporal dynamics, and stochastic variability. However, strong deterministic models already provide accurate initial reconstructions, leaving a structured residual error. We reformulate probabilistic imputation as a baseline-residual decomposition, where a pretrained model captures the dominant signal and a diffusion process models the residual uncertainty. Building on this idea, we propose RDDMPI, a conditional residual diffusion framework that operates directly in residual space. The model leverages both the deterministic baseline reconstruction and its latent representation as conditioning information, enabling targeted correction of missing values. This formulation reduces the effective complexity of the diffusion process, enabling it to focus on structured correction terms rather than reconstructing the full signal. Experiments on multiple benchmark datasets show that RDDMPI achieves state-of-the-art performance in both reconstruction accuracy and uncertainty quantification.