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

Area of Science:

• * Causal inference and machine learning for complex systems analysis.
• * Spatio-temporal data modeling and forecasting.
• * Development of interpretable AI for scientific discovery.

Background:

• * Multivariate spatio-temporal forecasting predicts interdependent variables across space and time.
• * Current methods struggle with dynamic causal dependencies and latent confounders.
• * Accurate causal modeling is crucial for interpretability, robustness, and decision support.

Purpose of the Study:

• * To propose MCST, a novel framework for comprehensive causal modeling in multivariate spatio-temporal forecasting.
• * To address limitations in capturing complete and dynamic causal relationships.
• * To enhance model interpretability, robustness, and predictive performance.

Main Methods:

• * Employing variational inference to disentangle exogenous factors and identify latent confounders.
• * Designing a causal estimator to quantify instantaneous and lagged causal transmissions across dimensions.
• * Integrating causal transmissions with Structural Causal Models (SCMs) for refined generation mechanisms.

Main Results:

• * MCST consistently outperforms existing methods in predictive performance across diverse datasets.
• * The framework provides enhanced interpretability through explicit causal reasoning.
• * Demonstrated effectiveness on three real-world and one synthetic dataset.

Conclusions:

• * MCST offers a significant advancement in multivariate spatio-temporal forecasting.
• * The framework successfully models dynamic causal dependencies and latent confounders.
• * MCST enhances both predictive accuracy and the interpretability of complex system models.