I-RCD: an improved algorithm of repetitive causal discovery from data with latent confounders

Discovering causal relationships from data affected by latent confounders is an important and difficult task. Until recently, approaches based on causal function models have not been used to present variable pairs whose relationships are affected by latent confounders, although some constraint-based methods are possible. Recently, a method based on causal function models called repetitive causal discovery (RCD), which infers causal relationships under the assumption that latent confounders exist, has been proposed. However, it has been pointed out that there are causal models RCD cannot identify. This problem is caused by the part of the RCD algorithm that extracts the set of ancestors of each observed variable. In this paper, we investigate the modifications to the RCD algorithm and propose an improved algorithm of RCD which we call improved RCD (I-RCD).The RCD algorithm removes the influence of the common ancestors of two variables from them when inferring the causal relationship between them, whereas the I-RCD removes the influence of the all ancestors of each variable from both variables respectively. The experimental results show that I-RCD accurately infers variable pairs with the same unobserved common causes and identify the direct causal relationships between observed variables compared to RCD. © 2022, The Behaviormetric Society.