2024. 1. 15. 08:00ㆍ코딩 도구/LG Aimers
LG Aimers: AI전문가과정 4차
Module 5. 『인과추론』
ㅇ 교수 : 서울대학교 이상학 교수
ㅇ 학습목표
본 모듈은 인과성에 대해 추론하고 경험적 데이터를 사용하여 인과 관계를 결정하는 방법을 익힘으로써 데이터를 생성한 프로세스에 대해
만들어야 하는 필수 가정과 이러한 가정이 합리적인지 평가하는 방법, 마지막으로 추정되는 양을 해석하는 방법에 대해 학습하고자 합니다.
Part 1. 인과성에 대한 소개 및 인과적 추론을 위한 기본 개념
-What is Causality?
Definition
“Causality ... is influence by which one event, process, state, or object (a cause)
contributes to the production of another event, process, state, or object (an
effect) where the cause is partly responsible for the effect, and the effect is
partly dependent on the cause.” (emphasis mine)*
인과성이란 하나의 사건, 과정, 상태 또는 대상(원인)에 의해 영향을 받는 것이다.
▶ (News) Papers: ‘increases’, ’decreases’ vs. linked to, associated with
▶ Daily Life: because, hence, thus, due to, ...
-Why Do We Study Causality?
Definition of Science Ņ
▶ “Knowledge or a system of knowledge covering
general truths or the operation of general laws
especially as obtained and tested through scientific method.”*
Causality in various academic disciplines
▶ Physics, Chemistry Ҏ, Biology, Climate Science װ,
▶ Psychology Ҙ, Social Science, Economics œ,
▶ Epidemiology, Public Health
(COVID-19, mask policy, social distancing, # of vaccination, side effects)
-How is Causality related to & {AI, ML, & DS}?
Artificial Intelligence
a rational agent performing actions to achieve a goal
e.g., reinforcement learning πθ (action | state)
Machine Learning
Currently focused on learning correlations,
e.g., Pˆθ (y|x) ≈ P(y|x)
Data Science
Capture, Process, Analyze (e.g., Stat, ML), Communicate with Data
-Pearl’s Causal Hierarchy
▶ Level 1: Associational or Observational
가장 기본적인 관측계층
->변수들의 상관성을 알 수 있다.
▶ Level 2: Interventional or Experimental
실험계층
->실험을 통해서 나오는 결과에 관심을 두는 것
▶ Level 3: Counterfactual
반사실적 계층
->관측값, 실험에 의한 값 동시 고려
-Simpson’s Paradox
Consider the following scenario:
1. A Patient with Kidney Stone Ҵ visits a Hospital.
2. A Doctor examines the Patient and provides a Treatment.
3. The Patient’s Health Outcome Ӊ is later reported
Healthcare Database!
-Lesson’s Learned from Simpson’s Paradox
▶ Causal analyses need to be guided by subject-matter knowledge .
▶ Identical data arising from different causal structures need to be analysed differently.
▶ No purely statistical rules exist to guide causal analyses.
-Data & Questions
-Causal Framework: Structural Causal Model
-Definition (Structural Causal Model)
A structural causal model (SCM) M is a 4-tuple ⟨V,U,F,P(U)⟩,
where
▶ U is a set of exogenous variables;
▶ P(U) is a distribution over U;
▶ V = {V1,...,Vn} are endogenous variables;
▶ F = { f1,... fn} are functions determining V,
vi ← fi(pai,ui)
where Pai ⊆ V\ {Vi}, Ui ⊆ U
-Intervention — do(·) operator
▶ Given a model M the action of fixing any observable variable X ∈ V to a constant value x is denoted using the do(·) operator as do(X = x).
▶ This operation gives birth to a submodel Mx that looks exactly like M but with functions where fx has been replaced with a constant x.
▶ These two graphs represent the world before and after an intervention do(X = x).
-Intervention — Causal Effects
Definition (Causal Effect)
Given two disjoint sets of variables, X and Y, the causal effect of X on Y, denoted as P(y|do(x)) or Px(y), is a function from X to thespace of probability d istributions of Y.
-Reading Conditional Independence from Causal Diagram
-Summary for Part 1
▶ Structural Causal Model 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 expertise.
▶ Intervention do(X = x) as a submodel Mx, which induces a manipulated causal graph GX.
▶ Causal effect of X = x on Y = y is defined as P(y|do(x)).
'코딩 도구 > LG Aimers' 카테고리의 다른 글
LG Aimers 4기 인과추론의 다양한 연구 방향 (0) | 2024.01.17 |
---|---|
LG Aimers 4기 인과추론 수행을 위한 기본 방법론 (0) | 2024.01.16 |
LG Aimers 4기 그리고 Ensemble Learning (0) | 2024.01.14 |
LG Aimers 4기 그리고 Advanced Classification Model (2) | 2024.01.13 |
LG Aimers 4기 그리고 Linear Classification (0) | 2024.01.12 |