Fast Causal Division for Supporting Robust Causal Discovery

Abstract

Discovering the causal relationship from the observational data is a key problem in many scientific research fields. However, it is not easy to detect the causal relationship by using general causal discovery methods among large scale data, due to the curse of the dimension. Although some causal dividing frameworks are proposed to alleviate these problems, they are, in fact, also faced with high dimensional problems. In this work, we propose a split-and-merge method for causal discovery. The original dataset is firstly divided into two smaller subsets by using low-order CI tests, and then the subsets are further divided into a set of smaller subsets. For each subset, we employ the existing causal learning method to discovery the corresponding structures, by combined all these structures, we finally obtain the complete causal structure. Various experiments are conducted to verify that compared with other methods, it returns more reliable results and has strong applicability.