Spatiotemporal Extrapolation Through Conditional Diffusion Probability Models

In recent years, spatiotemporal data have played a crucial role in weather, transportation, and disease transmission within the context of the Internet of Things (IoT). However, due to cost constraints and sensor failures, many regions lack observational data and remain unmonitored. This poses a challenge to the generalization of the model, which is often addressed by spatiotemporal completion methods. However, most existing interpolation and completion methods are limited to the data distribution of the training regions and struggle to generalize to out-of-distribution scenarios. This article addresses the challenge of generalization, particularly in scenarios that require inference from regions that have never been observed before. To overcome this limitation, we propose an inductive generative model for spatiotemporal extrapolation. Our approach is based on a denoising diffusion probabilistic model (DDPM), incorporating attention mechanisms guided by nonlocal features and dynamic topology information. This enables our model to generalize to previously unseen regions. Empirical evaluations of three datasets in real-world and cross-city evaluations demonstrate the superior performance of our approach over state-of-the-art methods.