基于CI测试的因果特征选择算法综述

Overview of causal feature selection algorithms based on CI testing

在机器学习和数据分析中,频繁出现的高维数据使得数据预处理变得愈发重要。作为预处理的重要方法,特征选择可以降低问题的复杂性。其中,传统的特征选择通过捕捉特征之间的相关性来实现,而基于因果关系的特征选择方法则更注重特征间的因果性。研究表明,因果特征选择具有更高的可解释性和鲁棒性,因此这一领域备受关注,在近20年来有许多算法被提出。文章首先回顾了这些算法,然后按照学习的策略对它们进行了分类,最后讨论并提出了因果特征选择中存在的问题和对应解决思路。

In machine learning and data analysis,the frequent occurrence of high-dimensional data makes data preprocessing increasingly important.As an important method of preprocessing,feature selection can reduce the complexity of the problem.Among them,traditional feature selection achieves it by capturing the correlation between features,while causal relationship based feature selection methods focus more on the causality between features.Research has shown that causal feature selection has higher interpretability and robustness,therefore this field has received great attention,and many algorithms have been proposed in the past two decades.This article first reviews these algorithms,then classifies them according to learning strategies,and finally discusses and proposes the problems and corresponding solutions in causal feature selection.