AdaST: Adaptive Coupling for Spatiotemporal Forecasting

Summary

AdaST is an adaptive framework for spatiotemporal forecasting, motivated by the observation that real-world data can follow different coupling regimes: temporal-dominated, spatial-dominated, or strongly coupled. The authors argue that methods assuming uniformly strong spatial-temporal coupling can learn spurious dependencies and lose accuracy when one type of correlation dominates. AdaST dynamically adjusts spatial and temporal modeling through a decompose-recompose design. It factorizes inputs into components representing different coupling patterns with heterogeneity-aware experts, processes each component with role-aligned modules, and combines them using a correlation-informed adaptive recomposer. The framework is designed to address unknown coupling structure, heterogeneous coupling dynamics, and weaknesses in spatial modeling. Extensive experiments are reported to show that AdaST significantly outperforms state-of-the-art baselines, supporting the value of adaptive coupling for spatiotemporal forecasting.