Learning General Causal Structures with Hidden Dynamic Process for Climate Analysis

The heart of climate analysis is a rational effort to understand the causes behind the purely observational data. Latent driving forces, such as atmospheric processes, play a critical role in temporal dynamics, and the task of inferring such latent forces is often a problem of Causal Representation Learning (CRL). Moreover, geographically nearby regions may directly interact with each other, and such direct causal relations among the observed data are often not modeled in traditional CRL, making the problem more challenging. In this paper, we propose a unified framework that can uncover not only the latent driving forces, but also the causal relations among the observed variables. We establish conditions under which the hidden dynamic process and the relations among the observed variables are simultaneously identifiable from time-series data. Even without parametric assumptions on the causal relations, we provide identifiability guarantees for recovering latent variables and the relations among the observed variables via contextual information. Guided by these insights, we propose a framework for nonparametric Causal Discovery and Representation learning (CaDRe), based on a time-series generative model with structural constraints. Synthetic data validates our theoretical claims. On real-world climate datasets, CaDRe achieves competitive forecasting performance and offers the visualized causal graphs consistent with domain knowledge, which is expected to improve our understanding of the climate systems.