Zhao Ruiqi, Yang Xiaoxia, He Yangbo
School of Mathematical Sciences, Peking University, Beijing 100871, China.
College of Science, Beijing Forestry University, Beijing 100083, China.
Entropy (Basel). 2024 Jun 24;26(7):541. doi: 10.3390/e26070541.
This paper addresses the challenge of identifying causes for functional dynamic targets, which are functions of various variables over time. We develop screening and local learning methods to learn the direct causes of the target, as well as all indirect causes up to a given distance. We first discuss the modeling of the functional dynamic target. Then, we propose a screening method to select the variables that are significantly correlated with the target. On this basis, we introduce an algorithm that combines screening and structural learning techniques to uncover the causal structure among the target and its causes. To tackle the distance effect, where long causal paths weaken correlation, we propose a local method to discover the direct causes of the target in these significant variables and further sequentially find all indirect causes up to a given distance. We show theoretically that our proposed methods can learn the causes correctly under some regular assumptions. Experiments based on synthetic data also show that the proposed methods perform well in learning the causes of the target.
本文探讨了识别功能动态目标成因的挑战，这些目标是随时间变化的各种变量的函数。我们开发了筛选和局部学习方法，以了解目标的直接成因以及给定距离内的所有间接成因。我们首先讨论功能动态目标的建模。然后，我们提出一种筛选方法来选择与目标显著相关的变量。在此基础上，我们引入一种结合筛选和结构学习技术的算法，以揭示目标及其成因之间的因果结构。为了解决长因果路径会削弱相关性的距离效应问题，我们提出一种局部方法来发现这些显著变量中目标的直接成因，并进一步依次找到给定距离内的所有间接成因。我们从理论上表明，在一些常规假设下，我们提出的方法能够正确地学习成因。基于合成数据的实验也表明，所提出的方法在学习目标成因方面表现良好。
