I-RCD: an improved algorithm of repetitive causal discovery from data with latent confounders
I-RCD: an improved algorithm of repetitive causal discovery from data with latent confounders
复制标题
DOI：
10.1007/s41237-022-00160-4
复制
发表时间：
2022-06
期刊：
影响因子：
--
通讯作者：
Takashi Nicholas Maeda
中科院分区：
文献类型：
--
作者：
Takashi Nicholas Maeda
从受潜在混杂因素影响的数据中发现因果关系是一项重要且困难的任务。直到最近，基于因果函数模型的方法还没有被用来呈现其关系受潜在混杂因素影响的变量对，尽管一些基于约束的方法是可能的。最近，提出了一种基于因果函数模型的方法，称为重复因果发现（RCD），该方法在存在潜在混杂因素的假设下推断因果关系。然而，有人指出，有些因果模型 RCD 无法识别。这个问题是由 RCD 算法中提取每个观察变量的祖先集合的部分引起的。在本文中，我们研究了对RCD算法的修改，并提出了一种改进的RCD算法，我们称之为改进RCD（I-RCD）。RCD算法在推断两个变量之间的因果关系时消除了两个变量的共同祖先的影响，而I-RCD则分别消除了两个变量中每个变量的所有祖先的影响。实验结果表明，与 RCD 相比，I-RCD 能够准确推断具有相同未观察到的共同原因的变量对，并识别观察到的变量之间的直接因果关系。
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.