ABSTRACT The analysis of spatiotemporal data is essential across many fields, such as transportation, meteorology and healthcare. Data gathered in practical applications often suffer from incompleteness due to device failures and network disruptions. Spatiotemporal imputation targets the estimation of missing observations by exploiting intrinsic spatial–temporal dependencies. Although traditional statistical and machine‐learning methods depend on restrictive distributional assumptions, graph‐ or recurrent‐based models accumulate errors through iterative propagation. Diffusion probabilistic models mitigate these issues by sampling directly from a learnt data prior instead of recycling past imputations. However, existing conditional diffusion variants still converge towards overly similar reconstructions, obscuring the genuine uncertainty and heterogeneity of real‐world traffic, environmental or clinical streams. Preserving—and faithfully quantifying—this intrinsic diversity is crucial for reliable forecasting and downstream decision‐making. We propose TSD, a conditional diffusion framework that integrates disentangled temporal representations and contrastive learning to improve generalisability in spatiotemporal imputation. Specifically, the approach uses disentangled temporal representations as conditional information to guide the reverse process. We also enhance the final loss using a contrastive learning strategy to improve representation quality, mitigating the impact of data missing completely at random (MCAR) and noise on learnt features. Through comprehensive experiments using three distinct real‐world datasets, TSD has competitive results compared to leading‐edge baselines.
Chen et al. (Mon,) studied this question.