๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
1๏ธโƒฃ AI•DS/๐ŸฅŽ Casual inference

์ธ๊ณผ์ถ”๋ก ์˜ ๋ฐ์ดํ„ฐ ๊ณผํ•™ - ๊ตฌ์กฐ์  ์ธ๊ณผ๋ชจํ˜•

by isdawell 2023. 5. 1.
728x90

์ฐธ๊ณ ์˜์ƒ : Bootcamp 5-3. ๊ตฌ์กฐ์  ์ธ๊ณผ๋ชจํ˜• 

 

 

 

 

 

1. Structural model


 

โ—ฏ  Causal inference = How to address endogeneity 

 

•  treatment ์— ๋Œ€ํ•œ selection process (data generation process) ๋ฅผ ์•Œ๊ธฐ ์–ด๋ ค์šด ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„์„œ, ์ธ๊ณผ์ถ”๋ก ์ด ์–ด๋ ค์›Œ ์ง€๋Š” ๊ฒƒ์ด๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ•๋ก ๋“ค์ด ๋“ฑ์žฅํ•œ ๊ฒƒ์ด๋‹ค. 

 

 

 

•  Design based approach : selection process ๋ฅผ ์•Œ ์ˆ˜ ์žˆ๋Š” research design ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ• 

•  Selection model : selection process ๋ฅผ ํ†ต๊ณ„์ ์œผ๋กœ ๋ชจ๋ธ๋ง ํ•จ์œผ๋กœ์จ selection process ์—์„œ selection bias ๋ฅผ ์ง์ ‘ ๊ณ„์‚ฐํ•˜๊ณ , ์ด๋ฅผ ์ง์ ‘ ํ†ต์ œํ•˜์ž๋Š” ์ ‘๊ทผ 

•  Causal graph : data generation process ๋ฅผ ํ†ต๊ณ„์ ์ธ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œํ˜„ํ•˜๊ณ , ๊ทธ ๊ทธ๋ž˜ํ”„๋ฅผ ํ™œ์šฉํ•ด์„œ data generation process ๋ฅผ ๊ฐ์•ˆํ•œ ์ƒํƒœ์—์„œ ์ธ๊ณผ์ถ”๋ก ์„ ํ•˜๊ณ ์ž ํ•˜๋Š” ์ ‘๊ทผ 

 

 

 

 

โ—ฏ  Structural causal model = probabilistic causal mechanisms 

 

•  Structural causal model : data generation ์„ ํ‘œํ˜„ํ•˜๋Š” causal mechanism ์ด ์กด์žฌํ•˜๊ณ  ์ด๊ฒƒ์„ ์ˆ˜ํ•™์ ์œผ๋กœ ํ‘œํ˜„ํ•œ ๋ชจ๋ธ 

 

 

•  Data generation process ๊ฐ€ structural causal model ๋กœ ์ •์˜ํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ณด์ง€๋งŒ,  ์‹ค์ œ ์„ธ๊ณ„์—์„œ๋Š” ์ด๋Ÿฌํ•œ causal model ์„ ๋ช…ํ™•ํžˆ ์ •์˜ํ•˜๊ธฐ๊ฐ€ ์‰ฝ์ง„ ์•Š๋‹ค. 

•  ์šฐ๋ฆฌ๊ฐ€ ์ถ”๋ก ํ•ด ๋ณผ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ํ™•๋ฅ  ๋ถ„ํฌ๋ฅผ ์ถ”๋ก ํ•ด ๋ณด๋Š” ๊ฒƒ 

 

 

•  ๊ด€์ฐฐ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ์™€ ๊ทธ์— ๋Œ€ํ•œ ํ™•๋ฅ ๋ถ„ํฌ์˜ 3๊ฐ€์ง€ ๊ตฌ๋ถ„ 

 

โ‘  ๊ด€์ฐฐ๋œ ๋ฐ์ดํ„ฐ์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋Š” ํ™•๋ฅ ๋ถ„ํฌ : associational distribution 

โ‘ก ์–ด๋–ค ๋ณ€์ˆ˜์— ๋Œ€ํ•œ ํ–‰๋™์„ ์ทจํ–ˆ์„ ๋•Œ ๋‚˜ํƒ€๋‚˜๋Š” ํ™•๋ฅ ๋ถ„ํฌ (์ธ๊ณผ๊ด€๊ณ„์— ๋Œ€ํ•œ ํ™•๋ฅ ๋ถ„ํฌ) : interventional distribution 

โ‘ข ํŠน์ • ํ–‰๋™์— ๋Œ€ํ•œ ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ๋„˜์–ด, ๋‹ค๋ฅธ ํ–‰๋™์„ ์ทจํ–ˆ๋‹ค๋ฉด ์–ด๋–ป๊ฒŒ ๋˜์—ˆ์„๊นŒ์— ๋Œ€ํ•œ ํ™•๋ฅ ๋ถ„ํฌ : counterfactual distribution 

 

โ†ช ์ƒ์œ„ ๋‹จ๊ณ„์˜ ํ™•๋ฅ ๋ถ„ํฌ๋Š” ์•„๋ž˜ ๋‹จ๊ณ„์˜ ์ •๋ณด๋งŒ์œผ๋กœ๋Š” ๋‹ตํ•  ์ˆ˜ ์—†๋‹ค. 

 

 

 

 

 

2. Causal inference with SCM 


 

โ—ฏ SCM 

 

•  causal inference ๋ผ๋Š” ๊ฒƒ์€, ๋ฐ์ดํ„ฐ์—์„œ ๊ด€์ฐฐ๋˜๋Š” ๋‹จ์ˆœํ•œ ํ™•๋ฅ ๋ถ„ํฌ๊ฐ€ ์•„๋‹ˆ๋ผ, ์–ด๋–ค ํ–‰๋™์„ ์ทจํ–ˆ์„ ๋•Œ ๋‚˜ํƒ€๋‚˜๋Š” ๊ฒฐ๊ณผ์— ๋Œ€ํ•œ ํ™•๋ฅ ๋ถ„ํฌ์ด๊ธฐ ๋•Œ๋ฌธ์—, Interventional distribution ๊ณผ counterfactual distribution ์„ ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ๊ฒฐ๊ตญ causal inference ๋‹ค. 

 

 

•  ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๊ฐ€ ์‹ค์ œ ์„ธ๊ณ„์—์„œ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์€ associational distribution ์ด๊ณ , ํ•˜์œ„๋‹จ๊ณ„์˜ ์ •๋ณด๋งŒ์„ ๊ฐ€์ง€๊ณ  ์ƒ์œ„๋‹จ๊ณ„์˜ ์ •๋ณด๋Š” ์•Œ ์ˆ˜ ์—†๋‹ค. ๋”ฐ๋ผ์„œ ๋ฐ์ดํ„ฐ ๊ทธ ์ด์ƒ์˜ ์ •๋ณด๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ๊ทธ๋ž˜์„œ causal graph ๋ฅผ ์ด์šฉํ•ด์•ผ ํ•œ๋‹ค. 

 

•  ๋น„๋ก data generation process ์— ๋Œ€ํ•œ ์‹ค์ œ causal SCM ์€ ์•Œ ์ˆ˜ ์—†์ง€๋งŒ, ์ ์–ด๋„ ๋ณ€์ˆ˜๋“ค์ด ์–ด๋– ํ•œ ๊ด€๊ณ„๋กœ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์„์ง€๋Š” ์•Œ ์ˆ˜ ์žˆ๋‹ค. Graphical model ์„ ํ™œ์šฉํ•ด์„œ ํ•˜์œ„๋‹จ๊ณ„ ์ •๋ณด๋กœ ์ƒ์œ„๋‹จ๊ณ„ ์ •๋ณด๋ฅผ ์ถ”์ •ํ•˜๊ณ ์ž ํ•˜๋Š” ๊ฒƒ์ด SCM ํ•˜์—์„œ์˜ causal inference ๋ผ๊ณ  ๋ณผ ์ˆ˜ ์žˆ๋‹ค. 

 

•  graphical model + associational distribution ๋ฅผ ํ†ตํ•ด interventional distribution, ๋” ๋‚˜์•„๊ฐ€ ์ด๋Ÿฐ ์ •๋ณด๋ฅผ ์ข…ํ•ฉํ•ด์„œ counterfactual distribution ์„ ๊ตฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋œ๋‹ค. 

 

 

 

 

3. Definition of Causal effect using do-operator 


 

โ—ฏ  do-operator 

 

•  SCM ์—์„œ๋Š” interventional distribution ์„ ์ด๋ก ์ ์œผ๋กœ ๊ตฌํ•˜๊ณ , associational distribution ๊ณผ causal graph ๋ฅผ ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ, ์‹ค์ œ๋กœ ๊ตฌํ•ด๋ณด์ž๋Š” ๊ฒƒ โ‡จ SCM ์—์„œ์˜ ์ธ๊ณผ์ถ”๋ก  : ์ด๋ก ์ ์œผ๋กœ ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ๋จผ์ € ์ •์˜ํ•˜๋Š” ๊ฒƒ์ด ํ•„์š”ํ•˜๋‹ค. 

 

•  do operator : ์ฃผ์–ด์ง„ causal graph ๋ฅผ ๊ทธ๋Œ€๋กœ ๋ถ„์„ํ•˜๋ฉด backdoor path ๋กœ ์ธํ•ด ์ƒ๊ด€๊ด€๊ณ„์™€ ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์—, ์ด๋ก ์ ์œผ๋กœ backdoor path ๊ฐ€ ์—†๋Š” ์ƒํƒœ์—์„œ ๊ณ„์‚ฐํ•œ ์ธ๊ณผ์  ํšจ๊ณผ๊ฐ€, ์šฐ๋ฆฌ๊ฐ€ ๊ณ„์‚ฐํ•ด์•ผ ํ•  ์ธ๊ณผ์  ํšจ๊ณผ๋ผ๊ณ  ์ •์˜ํ•˜๋Š” ๊ฒƒ 

 

 

•  ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” associational (conditional) distribution ์ธ  P(Y|X) ๋Š” ์ธ๊ณผ์  ํšจ๊ณผ๋ผ ๋งํ•  ์ˆ˜ ์—†๋‹ค. C ๋ผ๋Š” confounder ๋กœ ์ธํ•œ backdoor path ๊ฐ€ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. X๊ฐ€ Y์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ ๋ฟ ์•„๋‹ˆ๋ผ C๋กœ ์ธํ•œ non-causal association ๋„ ์žˆ๋‹ค.

 

 

•  ๋งŒ์•ฝ ์—ฌ๊ธฐ์„œ confounder ์—†์ด X๊ฐ€ ์‹ค์ œ ์›์ธ์ด์—ฌ์„œ ์ธ๊ณผํšจ๊ณผ๊ฐ€ ์žˆ๋‹ค๋ฉด ์–ด๋–จ๊นŒ โ‡จ ์ด๋ก ์ ์œผ๋กœ ๊ตฌํ•ด๋ด„ : do(X) : non causal effect ๋ฅผ ์ฐจ๋‹จํ•˜๋Š” ํ˜•์‹์œผ๋กœ ์ •์˜ํ•ด ๋ณผ ์ˆ˜ ์žˆ์Œ 

 

•  do() ํ•จ์ˆ˜๋Š” X์— ์–ด๋– ํ•œ action ์„ ์ทจํ•˜๋Š” ํ˜•ํƒœ๊ฐ€ ์•„๋‹ˆ๋ผ, treatment ๋ณ€์ˆ˜์˜ ๋ถ€๋ชจ๋…ธ๋“œ, ์ฆ‰ treatment ๋ณ€์ˆ˜์— ์˜ํ–ฅ์„ ์ฃผ๋Š” ๋‹ค๋ฅธ ์š”์ธ๋“ค์˜ ํšจ๊ณผ๋ฅผ ๋ฐฐ์ œํ•˜์ž๋Š” ์ด๋ก ์ ์ธ ๊ฐœ์ž…์ด๋‹ค.  interventional distribution 

 

 

•  ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด์„œ ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” conditional distribution ์„ ํ†ตํ•ด์„œ, ์ด๋ก ์ ์œผ๋กœ ๊ตฌ์ถ•ํ•œ interventional distribution ์„ ์–ด๋–ป๊ฒŒ ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š”์ง€๊ฐ€ ์ค‘์š” โ‡จ interventional distribution ์„ conditional distribution ์œผ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ๋Š”๊ฒŒ ์ค‘์š”ํ•˜๋‹ค. ๊ทธ๋Ÿฌํ•œ ๋ณ€ํ™˜์˜ ๊ณผ์ •์ด ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด identification ์ด๋ผ๊ณ  ๋ถ€๋ฅธ๋‹ค. ๋ถˆ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด non-identification ์ด๋ผ ๋ถ€๋ฅธ๋‹ค. 

 

 

 

 

โ—ฏ  do-calculus : ๊ทธ๋ ‡๊ฒŒ ๋ณ€ํ™˜ํ•˜๋Š”๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ๋Š” ์ผ๋ จ์˜ graphical rule ์„ do-calculus ๋ผ ์ •์˜

 

 

 

•  ์™ผ์ชฝ์€ do-operator ์—†์ด T๊ฐ€ Y ์— ๋ฏธ์น˜๋Š” ํšจ๊ณผ๋ฅผ ์‚ดํŽด๋ณด๋ฉด, ๋นจ๊ฐ„์ƒ‰ ์ ์„ ์ฒ˜๋Ÿผ non-causal path ๊ฐ€ ์กด์žฌํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ด๋Š” causal effect ๋ผ ๋ณผ ์ˆ˜ ์—†๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ก ์ ์œผ๋กœ ์ƒ๊ฐํ•ด์„œ do(T) ๋ฅผ ํ•œ๋‹ค๋ฉด, ๋นจ๊ฐ„์ƒ‰ path ๋Š” ๋ชจ๋‘ ์ฐจ๋‹จ๋˜๊ณ  ํ•ด๋‹น causal effect ๊ฐ€ ์ด๋ก ์ ์œผ๋กœ ๊ตฌํ•ด์•ผ ํ•  ์ธ๊ณผํšจ๊ณผ์— ํ•ด๋‹นํ•˜๊ฒŒ ๋  ๊ฒƒ์ด๋‹ค. 

 

•  ๊ทธ๋Ÿฌ๋‚˜ ์˜ค๋ฅธ์ชฝ์€ ์ด๋ก ์ ์œผ๋ก  ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ์ด์ง€๋งŒ, ์‹ค์ œ๋กœ ๊ตฌํ•˜๊ธฐ๋Š” ์–ด๋ ต๋‹ค โ‡จ ์˜ˆ์™ธ : Random assignment 

 

 

 

•  potential outcome framework ์—์„œ causal inference ์˜ gold standard ๋Š” Random assignment ์ธ๋ฐ, ์ด๋Š” SCM ์—์„œ๋„ ๋งˆ์ฐฌ๊ฐ€์ง€์ด๋‹ค. ๋ฐ”๋ผ๋ณด๋Š” ๊ด€์ ์ด ๋‹ค๋ฅธ ๊ฒƒ์ด๋‹ค.  potential outcome framework ์˜ ๊ฒฝ์šฐ์—๋Š” counterfactual ์„ ๊ฐ€์žฅ ์ž˜ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋Š” control group ์„ ์ฐพ๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ๋กœ ๋‘๊ณ  ์žˆ์–ด ์ด๋ฅผ random assignment ๋กœ ์‹คํ˜„ํ•ด๋ณด๊ฒ ์ž๋Š” ์ž…์žฅ์ธ ๊ฒƒ์ด๊ณ , SCM ๊ด€์ ์—์„œ๋Š” random assignment ๋Š” ์‹ค์ œ ๋ชจ๋ธ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ฃผ๋Š” ๋งค๊ฐœ์ฒด (๋žœ๋คํ•œ ๋ฐฐ์ •์„ ํ†ตํ•ด non-causal ํ•œ ํ™”์‚ดํ‘œ๋ฅผ ๋ชจ๋‘ ์‚ฌ๋ผ์ง€๊ฒŒ ํ•จ) ๋กœ ๋ณด๋Š” ๊ฒƒ์ด๋‹ค. 

 

 

 

 

 

 

4. Identification of causal effect 


 

โ—ฏ  Identification ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ• : backdoor criterion 

 

•  w2์™€ C๋ฅผ conditioning ํ•˜๋ฉด non-causal association ์ด ์ฐจ๋‹จ๋œ๋‹ค โ‡จ ์ด๋ก ์ ์œผ๋กœ ์ •์˜ํ•œ do operation ๊ณผ ๋™์ผํ•œ ํ˜•ํƒœ์˜ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๊ฒŒ ๋จ 

 

 

•  Backdoor criterion : ์–ด๋–ค ๋ณ€์ˆ˜๋ฅผ conditioning ํ•ด์•ผ์ง€ backdoor path ๋ฅผ ์ฐจ๋‹จํ•  ์ˆ˜ ์žˆ๋Š”์ง€ formal ํ•˜๊ฒŒ ์ •์˜ํ•œ ๊ฒƒ. backdoor criterion ์— ํ•ด๋‹นํ•˜๋Š” ๋ณ€์ˆ˜๋“ค์„ ๋ชจ๋‘ conditioning ํ•˜๋ฉด ์ด๋ก ์ ์œผ๋กœ ๊ณ„์‚ฐํ•œ intervention distribution ์„ ์‹ค์ œ๋กœ ๊ณ„์‚ฐ ๊ฐ€๋Šฅํ•œ association distribution ์œผ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ๊ณ , causal inference ๋ฅผ ํ•  ์ˆ˜ ์žˆ๋‹ค. 

 

 

 

 

โ—ฏ  do-calculus 

 

•  do-calculus : ๋ชจ๋“  ๊ทธ๋ž˜ํ”„์  ์ƒํ™ฉ์—์„œ ์ ์šฉ ๊ฐ€๋Šฅํ•œ identification ์„ ์œ„ํ•œ ๊ทธ๋ž˜ํ”„์  ๋ฒ•์น™ 

 

 

 

 

โ—ฏ  Identification of causal effect 

 

•  query : ์šฐ๋ฆฌ๊ฐ€ ๊ตฌํ•ด์•ผ ํ•˜๋Š” ๊ฒƒ. do operator ์— ๊ธฐ๋ฐ˜ํ•œ interventional distribution 

•  identification : causal graph ๋ฅผ ํ™œ์šฉํ•ด์„œ interventional distribution ์„ ์‹ค์ œ ๊ณ„์‚ฐ ๊ฐ€๋Šฅํ•œ ํ˜•ํƒœ์˜ ๋ถ„ํฌ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ 

 

 

 

 

โ—ฏ  Estimation of causal effect 

 

•   ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•ด์„œ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ, identification ๋‹ค์Œ ๋‹จ๊ณ„ 

 

 

โ€ป Dobule ML 

 

 

 

 

 

5.  Further reading and advanced topics 


 

 

 

•  SCM ์„ ์ž˜ ์„ค๋ช…ํ•œ ๋…ผ๋ฌธ! 

 

 

 

 

 

 

 

6.  Potential outcome framework   VS  Structural causal model 


 

โ—ฏ  ๋น„๊ต 

 

 

•  Potential outcome framework ๋Š” social science ์—์„œ ๋งŽ์ด ํ™œ์šฉ๋˜๊ณ  ์žˆ๊ณ , SCM ์€ ์ปดํ“จํ„ฐ ์‚ฌ์ด์–ธ์Šค์—์„œ ๋งŽ์ด ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค. 

 

 

 

โ—ฏ  ์ฐจ์ด์  (1) Manipulability 

 

•  ์›์ธ๋ณ€์ˆ˜์— ๋Œ€ํ•œ ์ฐจ์ด 

 

 

•  Potential outcome framework ๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ manipulability ๋ฅผ ๊ฐ€์ •ํ•˜๊ณ  ์žˆ๋‹ค. 

 

 

•  ๋ฐ˜๋ฉด SCM ์—์„œ manipulability ๋Š” ์•„๋ฌด๋Ÿฐ ์—ญํ• ์„ ํ•˜์ง€ ์•Š๋Š”๋‹ค. 

 

 

 

 

โ—ฏ  ์ฐจ์ด์  (2) Causal Structure/Knowledge 

 

•  ex. ์‚ฐ๋ชจ์˜ ์ƒํƒœ๊ฐ€ ์•„์ด์˜ ๊ฑด๊ฐ•์ƒํƒœ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ ๋ถ„์„ : causal graph ํ•˜์—์„œ causal knowledge ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ธ๊ณผ์ถ”๋ก ์ด ๊ฐ€๋Šฅ 

 

 

 

•  Potential outcome framework ์—์„œ๋Š” causal knowledge ๋ฅผ ๋ฐ˜๋“œ์‹œ ์š”๊ตฌํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋‹ค. ์™„๋ฒฝํ•œ ๊ตฌ์กฐ๋ฅผ ๊ฐ–์ถ”๊ณ  ์žˆ์ง€ ์•Š๋”๋ผ๋„, counterfactual ์— ๊ฐ€๊นŒ์šด ๋น„๊ต๊ฐ€๋Šฅํ•œ control group์„ ๊ณ ์•ˆํ•˜๋Š” ์ ์ ˆํ•œ ์—ฐ๊ตฌ๋””์ž์ธ๋งŒ์œผ๋กœ ์ธ๊ณผํšจ๊ณผ๋ฅผ ์ถ”๋ก  ๊ฐ€๋Šฅํ•˜๋‹ค. ๊ฐ€๋ น ์น ๋ ˆ ์ง€์ง„๋ฐœ์ƒ์‹œ ์˜ํ–ฅ์„ ๋ฐ›์•˜๋˜ ์ง€์—ญ์—์„œ ์ŠคํŠธ๋ ˆ์Šค๋ฅผ ๋ฐ›์•˜๋˜ ์‚ฐ๋ชจ๋“ค์„ treatment ๋กœ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š์•˜๋˜ ์ง€์—ญ์˜ ์‚ฐ๋ชจ๋“ค์„ control group ์œผ๋กœ ์„ค์ •ํ•˜์—ฌ ๋น„๊ตํ•ด ๋ณผ ์ˆ˜ ์žˆ๋‹ค. 

 

 

 

โ—ฏ  Policy-based  VS  Knowledge-based

 

 

 

 

 

 

 

 

728x90

๋Œ“๊ธ€