Diffusion-based Spatio-temporal Interpolation with Dynamic Sensor Sets

ICLR 2026 Conference Submission20219 Authors

Keywords: Diffusion, spatiotemporal, time-series, Fourier, transformer, imputation, interpolation, kriging

TL;DR: A novel diffusion-based approach to spatio-temporal interpolation surpasses traditional methods by adeptly multivariate incomplete data, long horizons and dynamic sensor setups.

Abstract: We tackle spatio-temporal interpolation for virtual sensors in sparse, partially observed, and dynamically changing networks. We introduce DynaSTI, a diffusion-based generative framework that is fully inductive to unseen locations, trains directly on incomplete observations, and remains effective without retraining when sensor networks change with time. Our contributions are threefold: (i) a unified conditioning strategy that yields calibrated predictive distributions and robust performance under severe input-sensor dropout; (ii) a Fourier-domain compression variant, FDynaSTI, that accelerates sampling performance, and (iii) state-of-the-art performance on multiple real-world datasets, improving both RMSE and CRPS relative to strong baselines. Together, these results establish diffusion-based, frequency-aware probabilistic interpolation as a scalable solution for real-world, dynamic sensor networks.

Primary Area: generative models

Submission Number: 20219