因果推断学习笔记(五):画画因果图
很早很早就想要学习一下怎么画因果图,最近终于有时间折腾一下啦!这两周研究了一下怎么用非实验数据画因果图。关于因果图的理论知识这里就不再赘述,想要快速学习推荐读 A Survey of Learning Causality with Data: Problems and Methods [1] 的第四章,想要深度学习可以阅读 Judea Pearl 大佬的 Causality 教材 [2]。这篇文章主要是一些开源工具的使用体验。
经典算法介绍
使用非实验数据画因果图的方法主要可以分为以下几类 [2]:
- Constraint-based Algorithms. “This class of algorithms learn a set of causal graphs which satisfy
the conditional independence embedded in the data. These algorithms use statistical tests to verify
if a candidate graph fits all the independence based on the faithfulness assumption.”- 常见算法:PC 和变种、FCI、GFCI、RFCI……
- Score-based Algorithms. “Score-based algorithms replace conditional independence tests with the goodness of fit tests. Score-based algorithms learn causal graphs by maximizing the scoring criterion which returns the score of the causal graph given data.”
- 常见算法:Greedy Equivalence Search (GES)、FGES……
- Algorithms based on Functional Causal Models (FCMs). “In FCMs, a variable can be written as a function of its directed causes and some noise term.”
- 常见算法:Linear Non-Gaussian Acyclic Model (LiNGAM)、ICA-LiNGAM……