Can We Evaluate Causal Structures Using Graph Neural Networks?

Causal models have significant potential to augment prediction models through improved interpretability and regularisation. However, their applicability is limited as they are computationally expensive, complex, and hard to verify particularly if dealing with large-scale, heterogeneous, non-stationary datasets. In this paper we present a framework for improving the accessibility of causal tools for large scale datasets through an analysis of decomposition and subsampling methods that we evaluate on the popular causal discovery method PCMCI+. Further, we propose a novel method for causal structure evaluation that utilises the regularisation effect of causal modelling to evaluate candidate causal structures on data without the need of ground-truth.