Causality文献总结

今天就发表我的一篇一篇文献笔记，我目前的研究方向是Causality，主要应用在解耦学习中，另外还突发奇想发表了一篇关于Causality在feature importance的应用的文章在arxiv上：https://arxiv.org/abs/2308.14474

On Pearl’s Hierarchy and the Foundations of Causal  Inference, Elias Bareinboim, Columbia University, 2021.3

一. 阅读目的

精读—— 整理Judea工作，causal inference的脉络和参考文献

二. 文献内容

1. Introduction：

(1)长期目标：By breaking down a phenomenon into modular components, and describing how they interact to produce an emergent behavior or a final product, scientists seek to uncover the underlying data-generating processes, or features thereof.

(2) 短期目标：cross-layer inference through causal inference

(3) 发展：observable traces——data

(4) 前景：类比computation, an intuitive, ordinary concept underwent mathematical formalization, before then entering engineering practice.

(5 )两个问题待解决：                                                                                                             (1)The a. causal mechanisms underlying the phenomenon under investigation should be accounted for – indeed, formalized – in the analysis.   （SCM）                                           

b. This collection of mechanisms (even if mostly unobservable) should be formally tied to its output: the generated phenomena and corresponding datasets. （PCH）

2. Structural Causal Models and the Causal Hierarchy

(1)Causal Hierarchy:

    Associational -----> Interventional -----> Counterfactual

(2) SCM

 (3) Syntex

• L1-----seeing

• L2-----doing

• L3-----imaging

PNS: PNS encodes the extent to which a certain treatment to a particular outcome would be both necessary and sufficient

the probability of sufficiency:

the probability of necessity :

 3.  A Logical Perspective

 4. A Graphical Perspective

Object：causal inference——“climbing up” the PCH via lower-level data( the conditions under which valid causal claims can be made )

(1)Markovian graph

• Causal Inference via L1 -constraints：BN and conditional independence

• Causal Inference via L2 -constraints

Tip: G has an arrow from A to B (A → B) if B “listens” to the value of A

 

a. every variable is independent of its non-descendants given its parents

b. after fixing the parents of a variable by intervention, the corresponding function should be insensitive to any other intervention elsewhere in the system

c. Whether the function fi takes the value of its arguments following an intervention or by observation, the same behavior for Vi is observed 

• 由L1 data-----> L2 knowledge

2. Semi-markovian（confounded）: 外生变量互相依赖

• 外生变量依赖的图形语言

• 半马尔科夫模型的因果图

• 半马尔科夫模型的CBN

三. 总结：

本文系统概述了近期因果科学的发展，由于处于启蒙期，因此着重于着重于因果语言系统的搭建以及阐述部分目前推导出的定理引理，可以作为因果AI的“字典”使用。