Enhancing multivariate spatio-temporal forecasting via complete dynamic causal modeling.

Abstract

Multivariate spatio-temporal forecasting aims to predict the future evolution of multiple interdependent variables distributed across space and time. Effectively capturing the underlying causal dependencies among these variables is essential for enhancing model interpretability, robustness, and decision support in complex systems. However, existing methods often fall short in modeling complete and dynamic causal dependencies due to the presence of latent confounders and the challenges of identifying multidimensional causal interactions. To address these challenges, we propose MCST, a novel framework that systematically refines the causal generation process of each variable through comprehensive causal modeling. MCST first applies variational inference to disentangle variable-specific exogenous factors and identify latent confounders within a shared latent space. To capture dynamic causal dependencies, we design a causal estimator that quantifies both instantaneous and lagged causal transmission across spatial, temporal, and inter-variable dimensions. These estimated causal transmissions are then integrated with exogenous and endogenous components using SCMs, enabling the construction of refined, variable-wise causal generation mechanisms for accurate forecasting. Extensive experiments on three real-world and one synthetic dataset demonstrate that MCST consistently outperforms existing approaches in predictive performance while providing enhanced interpretability through explicit causal reasoning.