Post

인과와 인과추론 | Causality and Causal inference


인과, 인과추론의 개념과 관련 이론 (Back-door, Do-calculus) 들을 알아봅시다.

구조적 인과 모델(SCM)은 관측과 개입을 통해 인과 관계를 설명하는데 도움을 줍니다.

Back-door 기준은 인과 관계를 확인하고 혼동을 줄이는 데 도움이 되며, Do-계산법은 확률을 처리하는 데 유용한 규칙을 제공합니다.

     

Causality

  • Influence by shich one event, process, state, or object a contributes to the production of another event, process, state, or object where the cause is partly responsible for the effect, and the effect is partly dependent on the cause.
  • Causality in various academic disciplines
    • Physics, chemistry,biology, climate science
    • Psychology, social science, economics
    • Epidemiology, public health
  • Relation to AI, ML, DS
    • AI : a rational agent performing actions to achieve a goal (reinforcement learning)
    • ML : currently focused on learning correlations
    • DS : capture, process, analyze, communicate with data

     

Structural Causal Model (SCM)

  • SCM \(M = <U,V,F,P(U)>\) provides a formal framework.
  • SCM induces observational, interventional, and counterfactual distributions.
  • SCM induces a causal graph \(g\), which implies conditional independencies testable via d-separation (blockage).
  • The underlying model \(M\) is unknown but the causal graph \(g\) can be given from common sense or domain knowledge.
  • Intervention do(X=x) as a submodel Mx, which induces a manipulated causal graph \(g_\bar{x}\).
  • Causal effect of \(X=x\) on \(Y=y\) is defined as \(P(y\mid{do(x)})\).

Remark

  • Identifiability : causal effect may be computable from existing observational data for some causal graphs.
  • In a Markovian case an singleton X, a causal effect can be easily derivable by canceling output \(P(x\mid{pa_x})\)

     

Back-door Criterion

  • DefinitionBack-door

    • Find a set \(Z\) s.t. it can sufficiently explain ‘confounding’ between \(X\) and \(Y\). Then,
    \[P(y|do(x))=\sum_Z{P(y|x,z)P(z)}\]
  • DefinitionㅣBack-door criterion

    • A set \(Z\) satisfies the back-door criterion with respect to a pair of variables \(X, Y\) in causal diagram \(g\) if;
      • (i) no node in \(Z\) is a descendant of \(X\); and
      • (ii) $Z$ blocks every path between X ∈ \(X\) and Y ∈ \(Y\) that contains an arrow into X.
  • A back-door adjustment formula is simple and widely used but limited.

Back-door sets as substitutes of the direct parents of X

  • Rain satisfies the back-door criterion relative to Sprinkler ans Wet:
    • (i) Rain is not descendant of Sprinkler, and
    • (ii) Rain blocks the only back-door path from Sprinkler to Wet.
  • Adjusting for the direct parents of Sprinkler, we have:
\[P(\text{wt}|do(\text{sp}))=\sum_\text{sn}P(\text{wt}|\text{sp,sn})P(\text{sn})=\cdots=\sum_\text{rn}P(\text{wt}|\text{sp,rn})P(\text{rn})\]

     

Rules of Do-calculus

  • Backdoor criterion results in a very specific form of indentification formula.

  • Do-calculus (Pearl, 1995) provides general machinery to manipulate observational and interventional distributions.

  • TheoremㅣRules of Do-calculus (simplified)

    • Rule 1 : Adding/removing observations
    \[P(y|do(x),z)=P(y|do(x))\,\,\,\text{if}\,\,(Z\perp{Y|X})\,\,\text{in}\,\,g_{\bar{X}}\]
    • Rule 2 : Action/observation exchange
    \[P(y|do(x),do(z))=P(y|do(x),z)\,\,\,\text{if}\,\,(Z\perp{Y|X})\,\,\text{in}\,\,g_{\bar{X}\underline{Z}}\]
    • Rule 3 : Adding/removing actions
    \[P(y|do(x),do(z))=P(y|do(x))\,\,\,\text{if}\,\,(Z\perp{Y|X})\,\,\text{in}\,\,g_{\bar{X}\bar{Z}}\]
  • Do-calculus is sound and complete but it has no algorithmic insight
  • A graphical condition and an efficient algorithmic procedure have developed for identifiability.

  • Do-calculus is a set of rules to manipulate observational or interventional probabilites. (Do-calculus is complete)

     

Modern Identification Tasks

  • Experimental conditions ➔ Generalized identification

    • Combining datasets of different experimental conditions

    • The identifiability of any expression of the form \(P(y\mid{do(x), z})\) can be determined given any causal graph \(g\) and an arbitrary combination of observational and experimental studies.
    • If the query is identifiable, then its estimand can be derived in polynomial time.
  • Environmental conditions ➔ Transportability

    • Combining datasets from different sources

    • Non-parametric transportability can be determined provided that the problem instance is encoded in selection diagrams.
    • When transportability is feasible, the transport formula can be derived in polynomial time.
    • The causal calculus and the corresponding transportation algorithm are complete.
  • Sampling conditons ➔ Recovering from selection bias

    • Nonparametric recoverability of selection bias from causal and statistical settings can be determined provided that an augmented causal graph is available.
    • When recoverability is feasible, the estimated can be derived in polynomial time.
    • The result is complete for pure recoverability, and sufficient for recoverability with external information.
  • Responding conditons ➔ Recovering from missingness

     

Reference

본 포스팅은 LG Aimers 프로그램에서 학습한 내용을 기반으로 작성된것입니다. (전체 내용 X)

  1. LG Aimers AI Essential Course Module 5. 인과추론, 서울대학교 이상학 교수
This post is licensed under CC BY 4.0 by the author.