STCTN：一种基于时域偏倚校正与空域因果传递的时空因果表示学习方法

Spatio-temporal data, sampled from complex dynamical systems, is ubiquitous in real world, e. g., traffic flow, meteorological records and energy consumption. Learning effective feature representation from spatiotemporal data is the foundation of spatiotemporal data mining. The existing models overemphasize the statistical correlations in the spatiotemporal data and are susceptible to spurious correlations, which makes it hard to extract unbiased and robust feature representation. We model the generation process of spatiotemporal data based on the structural causal model （SCM）, analyze the causes of spurious correlations in observations, and propose a spatiotemporal causal representation learning method based on temporal bias adjustment and spatial causal transition. Here, we focus on two challenges in spatio-temporal representation learning: （1） Eliminating temporal confounding bias. Existing models fail to handle causal relationships and certainly not eliminate the influence of confounders in temporal domain. Hence, the second challenge is to remove confounding bias and extract unbiased temporal representations. （2） Modelling spatial causal relationships. Limited by predefined graph structures, existing models are susceptible to non-causal spatial spurious correlations so it is significant to recover underlying spatial causal structure under causal constraints. First, we eliminate temporal spurious correlation based on backdoor adjustment. Then, we construct causal transition network for eliminating spatial spurious correlation. Finally, the downstream feature decoder applies the causal representation to downstream tasks. We tackle spatio-temporal representation learning tasks from a causal perspective and analyze the causes of spatial and temporal spurious correlation in observation data. To the best of our knowledge, this is the first attempt to apply causal theory to spatio-temporal representation learning. Experiments on two real-world datasets show that the proposed spatiotemporal representation learning method effectively avoids the interference of spurious correlations, enhances the stability of the model, and reduces the prediction error of nodes with sparse data in two downstream prediction tasks by 3% and 10% respectively.