【AREA-introduction】Causal Science

Posted on Edited on In Views: Word count in article: 289 Reading time ≈ 1 mins.

causal AI learning (building)

Background knowledges

基于约束的方法

Graph

  • 3 kinds:

  • chain X->Z->Y

  • fork X<-Z->Y
    Z是X和Y的共因，如果X发生改变，Y很有可能也应该改变
    when the middle Z is given,X and Y won’t be affected anymore.

  • collider x->Z<-Y
    在没有given Z时，X和Y是独立的
    when the middle Z is given,X/Y would need to change to make sure Z won’t change while Y/X is changing

  • Path can’t be blocked by S(node set).

  • blocked:

    • given Z when chain/fork
    • not given Z(and its son nodes) when collider

  • D-Seperation:
    when every path between X and Y is blocked by Z

  • Markovian assumption

  • Faithfulness assumption
    由于以上两个假设我们可以通过given数据推出对应的Markov等价类

典型算法

IC、PC、FCI、RFCI、GFCI
基本思想是1.通过数据找到条件独立性列表(cond. independences) 2.通过条件独立性列表得到DAGs有向无环图

Markov Equivalence Class马尔科夫等价类

  • 可以由同一组d-seperation得到的图s
  • 马尔科夫等价类中的图有相同的骨架skeleton和V-结构
    比如chain和fork的3中图就有相同的V结构

基于分数的方法 score-based

Causal inference