Causal Inference: Question of Why?
ยท
Data Science/Econometrics
๋ณธ ํฌ์ŠคํŒ…์€ 2023. 4. 10. ๊ฒŒ์‹œ๋œ Youtube - ๊น€์„ฑ๋ฒ” [๊ต์ˆ˜/์‚ฐ์—…๊ฒฝ์˜๊ณตํ•™๋ถ€] - [Open DMQA Seminar] Causal Inference์˜ lecture note๋ฅผ ๋‹ค๋ฃน๋‹ˆ๋‹ค.https://www.youtube.com/watch?v=mxRhFYrQROU ์ธ๊ณผ๊ด€๊ณ„, ์ธ๊ณผ๊ด€๊ณ„์™€ ์ƒ๊ด€๊ด€๊ณ„๋Š” ๋‹ค๋ฅธ ํฌ์ŠคํŒ…์—์„œ ์ถฉ๋ถ„ํžˆ ๋‹ค๋ฃจ๊ณ  ์žˆ์œผ๋ฏ€๋กœ ๊ฐ„๋žตํžˆ ๋„˜์–ด๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค.์ธ๊ณผ๊ด€๊ณ„๋ž€?์–ด๋–ค ์‚ฌ๊ฑด์ด ๋‹ค๋ฅธ ์‚ฌ๊ฑด์˜ ์›์ธ์ด ๋˜์—ˆ๋‹ค๊ณ  ์ถ”๋ก ํ•˜๋Š” ๊ฒƒ (์›์ธ๊ณผ ๊ฒฐ๊ณผ ๊ด€๊ณ„) ์ƒ๊ด€๊ด€๊ณ„์™€ ์ธ๊ณผ๊ด€๊ณ„์ƒ๊ด€๊ด€๊ณ„๋Š” ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ์˜๋ฏธํ•˜์ง€ ์•Š์Œ์ƒ๊ด€๊ด€๊ณ„ (Correlation): ํ†ต๊ณ„์  ๋ณ€์ˆ˜์™€ ๋‹ค๋ฅธ ํ†ต๊ณ„์  ๋ณ€์ˆ˜๋“ค์ด ๊ณต๋ณ€(covariance)ํ•˜๋Š” ๊ด€๊ณ„ => ์˜ˆ์ธก ๋ชฉ์ ์ธ๊ณผ๊ด€๊ณ„ (Causality): ์„ ํ–‰ํ•˜๋Š” ํ•œ ๋ณ€์ˆ˜๊ฐ€ ํ›„ํ–‰ํ•˜๋Š” ๋‹ค๋ฅธ ๋ณ€์ˆ˜์˜ ์›์ธ์ด..
[Econometrics, ๊ณ„๋Ÿ‰๊ฒฝ์ œํ•™] SUTVA & ATU, ATE, ATT
ยท
Data Science/Econometrics
์ฐธ๊ณ ์ž๋ฃŒ: https://www.youtube.com/watch?v=e2BzsceHNX8 SUTVA (Stable Unit Treatment Value Assumption)"์•ˆ์ •๋œ ๋‹จ์œ„ ์ฒ˜๋ฆฌ ๊ฐ’ ๊ฐ€์ •", ์ฃผ๋กœ treatment๊ฐ€ ์•ˆ์ •์ ์œผ๋กœ ์ž˜ ์ด๋ค„์งˆ ์ˆ˜ ์žˆ๋Š”์ง€๋ฅผ ๋‹ค๋ฃจ๋Š” ๊ฐ€์ •์ž„.Potential Outcome Framework๊ฐ€ ์„ฑ๋ฆฝ๋˜๊ธฐ ์œ„ํ•œ ๊ธฐ๋ณธ ๊ฐ€์ •์ด์ž ์ธ๊ณผ์ถ”๋ก ์—์„œ ๊ฐ€์žฅ ๊ธฐ๋ณธ์ด ๋˜๋Š” ์ „์ œ๋‘ ๊ฐ€์ง€ ์กฐ๊ฑด์œผ๋กœ ๊ตฌ์„ฑ๋จ Casual effect= potential outcome with treatment - potential outcome without treatment("์ฒ˜์น˜ ๊ทธ๋ฃน์— ๋Œ€ํ•ด์„œ" ์ฒ˜์น˜๊ฐ€ ์žˆ์„ ๋•Œ ์ž ์žฌ์  ๊ฒฐ๊ณผ - ์ฒ˜์น˜๊ฐ€ ์—†์„ ๋•Œ ์ž ์žฌ์  ๊ฒฐ๊ณผ)=> observed outcome with treatme..
[Econometrics, ๊ณ„๋Ÿ‰๊ฒฝ์ œํ•™] ๊ณ ์ • ํšจ๊ณผ(Fixed Effect) ์‚ฌ๋ก€ ๋ถ„์„
ยท
Data Science/Econometrics
https://www.youtube.com/watch?v=UNIF-jHY2Rk๋“ค์–ด๊ฐ€๊ธฐ ์ „์—https://heesleisure.tistory.com/55 [Econometrics, ๊ณ„๋Ÿ‰๊ฒฝ์ œํ•™] ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ์™€ ๊ณ ์ • ํšจ๊ณผ(Fixed Effect)https://www.youtube.com/watch?v=yZOTgQfVfxQ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐCross-sectional Data : ์—ฌ๋Ÿฌ ์œ ๋‹›, ํŠน์ • ์‹œ์ ์˜ ์ •๋ณด๋ฅผ ํฌํ•จ (์„œ๋ฒ ์ด ๋“ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ์Œ)Time-Series Data : ํ•˜๋‚˜์˜ ์œ ๋‹›, ์‹œ๊ฐ„์— ๋”ฐ๋ฅธ ๋ณ€ํ™” (์ฃผ์‹ ๊ฐ€๊ฒฉ, ๊ตญ๊ฐ€์˜ Gheesleisure.tistory.com  ์ด์ „ ๊ธ€์„ ๋จผ์ € ์ฝ์–ด๋ณด์‹œ๋ฉด ์ดํ•ด๊ฐ€ ๋น ๋ฅด์‹ค ๊ฒ๋‹ˆ๋‹ค! How Fixed Effects Work for Causal Inference์–ด๋–ค ๋ฐฉ์‹์œผ๋กœ ์ž‘..
[Econometrics, ๊ณ„๋Ÿ‰๊ฒฝ์ œํ•™] ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ์™€ ๊ณ ์ • ํšจ๊ณผ(Fixed Effect)
ยท
Data Science/Econometrics
https://www.youtube.com/watch?v=yZOTgQfVfxQ ๋ฐ์ดํ„ฐ ๊ตฌ์กฐCross-sectional Data : ์—ฌ๋Ÿฌ ์œ ๋‹›, ํŠน์ • ์‹œ์ ์˜ ์ •๋ณด๋ฅผ ํฌํ•จ (์„œ๋ฒ ์ด ๋“ฑ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ์Œ)Time-Series Data : ํ•˜๋‚˜์˜ ์œ ๋‹›, ์‹œ๊ฐ„์— ๋”ฐ๋ฅธ ๋ณ€ํ™” (์ฃผ์‹ ๊ฐ€๊ฒฉ, ๊ตญ๊ฐ€์˜ GDP ๋“ฑ)Panel (longitudinal) Data : ์—ฌ๋Ÿฌ ์œ ๋‹›, ์‹œ๊ฐ„์— ๋”ฐ๋ฅธ ๋ณ€ํ™”๋™์ผํ•œ ์œ ๋‹›์— ๋Œ€ํ•ด ์‹œ๊ฐ„์— ๋”ฐ๋ฅธ ๋ณ€ํ™”๋ฅผ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•จ์ฆ‰, ๊ฐ ์‹œ์ ๋งˆ๋‹ค ์œ ๋‹›์ด ๋‹ค๋ฅด๊ฑฐ๋‚˜ ์œ ๋‹› ID๊ฐ€ ์—†์–ด์„œ ์ถ”์ ์ด ๋ถˆ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด? pooled cross-sectional data ๋จ  Recall How to Remove the Selection Bias ์ธ๊ณผ์ถ”๋ก ์„ ๋ฐฉํ•ดํ•˜๋Š” ์ตœ๋Œ€์˜ ์ ์€? ์„ ํƒ ํŽธํ–ฅ (Treatment๊ฐ€ ์—†๋Š” ์ƒํ™ฉ์—์„œ..
Ensemble(์•™์ƒ๋ธ”): Bagging๊ณผ Boosting, Stacking
ยท
Data Science/ML
Ensemble(์•™์ƒ๋ธ”)์ด๋ž€?์—ฌ๋Ÿฌ๊ฐœ์˜ ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ทธ ์˜ˆ์ธก์„ ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ ์ตœ์ข… ์˜ˆ์ธก์„ ๋„์ถœํ•˜๋Š” ๋ฐฉ๋ฒ• Bagging๊ณผ Boosting์˜ ์ฐจ์ด์ Bagging: ๋ณ‘๋ ฌ์ , parallelํ•จBootstrap Aggregating, ๋ถ€ํŠธ์ŠคํŠธ๋žฉ(๋žœ๋ค ์ƒ˜ํ”Œ๋ง)์„ ์ง‘๊ณ„ํ•˜๋Š” ๊ฒƒ, ์—ฌ๋Ÿฌ ์ƒ˜ํ”Œ ์กฐํ•ฉ์— ๋Œ€ํ•˜์—ฌ ํ•˜๋‚˜์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜(๋ชจ๋ธ) ํ™œ์šฉ์ƒ˜ํ”Œ๋ง ๊ณผ์ •์€ ๋ณต์›์ถ”์ถœ์ž„๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ์ง€ ์•Š์•„๋„ ์ถฉ๋ถ„ํ•œ ํ•™์Šตํšจ๊ณผ๋ฅผ ์ฃผ์–ด underfitting, overfitting ๋ฌธ์ œ์— ๋„์›€๋จBagging์˜ ๋Œ€ํ‘œ์ ์ธ ๊ธฐ๋ฒ•์œผ๋กœ RandomForest๊ฐ€ ์žˆ์Œ  Boosting: ์ˆœ์ฐจ์ , sequentialํ•จ์˜ค๋ถ„๋ฅ˜๋œ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด์„œ ๊ฐ€์ค‘์น˜๋ฅผ ์ฃผ๊ณ  ๋‹ค์Œ ๋ชจ๋ธ(๋ถ„๋ฅ˜๊ธฐ)์—์„œ ํ•ด๋‹น ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•˜๊ฒŒ ์œ ๋„๋จผ์ € ์ƒ์„ฑ๋œ ๋ชจ๋ธ์„ ์ ์  ๋ฐœ์ „์‹œ์ผœ๋‚˜๊ฐ€๋Š” ๋ฐฉ๋ฒ•์ž„   Stac..
[MODUPOP Seminar] Gen AI Safety & Security
ยท
Data Science/LLM
https://festa.io/events/4739 [๊ฐ•๋‚จ] ๋ชจ๋‘ํŒ ์„ธ๋ฏธ๋‚˜ | Gen AI Safety & Security | Festa! Festa์—์„œ ๋‹น์‹ ์ด ์ฐพ๋Š” ์ด๋ฒคํŠธ๋ฅผ ๋งŒ๋‚˜๋ณด์„ธ์š”. festa.io ์ด๋ฒˆ ์ฃผ์ œ๋Š” ๋ชจ๋‘์˜์—ฐ๊ตฌ์†Œ์—์„œ ์ฃผ์ตœํ•œ [Gen AI Safety & Security] ์˜คํ”ˆ ์„ธ๋ฏธ๋‚˜์ด๋‹ค. ๋ชจ๋‘์ฝ˜ ์ดํ›„๋กœ ํ”Œ์นœ์ด ๋œ ๋‚˜์™€ ๋ชจ๋‘์—ฐ.. ์š”์ƒˆ ๊ฐœ์ธ์—ฐ๊ตฌ&ํ”„๋กœ์ ํŠธ๋กœ LLM๊ณผ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•œ์ฐฝ ๊ณต๋ถ€ํ•˜๊ณ  ์žˆ์—ˆ๋Š”๋ฐ ์ด๋Ÿฐ ์„ธ๋ฏธ๋‚˜๋ผ๋‹ˆ! ๊ทผ๋ฐ ์‹ ์ดŒ๊ณผ ์ •๋ฐ˜๋Œ€ ๊ฐ•๋‚จ์บ ํผ์Šค๋ฅผ ์ €์‹œ๊ฐ„์— ๊ฐ€๋Š”๊ฑด ๋ฌด๋ฆฌ๋ผ ์˜จ๋ผ์ธ ์ˆ˜๊ฐ•์„ ํ•˜๊ธฐ๋กœ ํ–ˆ๋‹ค. (์ง€๊ธˆ ์ƒ๊ฐํ•˜๋‹ˆ ์‹ ์˜ ํ•œ ์ˆ˜์ž„) ์—ฐ๊ตฌ์‹ค ๋ฉค๋ฒ„๋“ค์—๊ฒŒ ๊ฐ™์ด๋“ฃ์ž๊ณ  ๊ผฌ์‹œ๋Š”์ค‘ ์ผ๋‹จ์€ ๋‚˜๋งŒ ๋“ค์€ ๊ฒƒ ๊ฐ™๊ธดํ•ด ... ์•„์‰ฌ์›€์„ ๋’ค๋กœํ•˜๊ณ ,, ์˜ค๋Š˜ ์ฃผ์ œ๋Š” ํœ˜๋ฐœ์„ฑ์ด ๊ฝค ๊ฐ•ํ•œ ๋‚ด์šฉ์ด๋ผ ๊ฐ„๋žตํžˆ ์ •๋ฆฌ๋งŒ ํ•ด๋ณด๋ คํ•œ๋‹ค ๋ชฉ์ฐจ..
[1๊ฐ•][CV] ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ์ดˆ ์ด๋ก  CS231n
ยท
Data Science/CS231n
๋ณดํ˜ธ๋˜์–ด ์žˆ๋Š” ๊ธ€์ž…๋‹ˆ๋‹ค.
[LLM] LangChain์œผ๋กœ ChatGPT API ํ™œ์šฉ & ํ”„๋กฌํ”„ํŠธ ํ…œํ”Œ๋ฆฟ
ยท
Data Science/LLM
์ถœ์ฒ˜: https://www.youtube.com/watch?v=BLM3KDaOTJM&list=PLQIgLu3Wf-q_Ne8vv-ZXuJ4mztHJaQb_v&index=3 https://www.youtube.com/watch?v=y6D5Hn_k4lE&list=PLQIgLu3Wf-q_Ne8vv-ZXuJ4mztHJaQb_v&index=5 API๋ฅผ ํ†ตํ•ด ChatGPT์™€ ๋Œ€ํ™”ํ•˜๊ธฐ davinci-003 vs gpt-3.5-turbo: GPT ๋ฒ„์ „ ์ฐจ์ด(3๊ณผ 3.5) temperature: chat-gpt๊ฐ€ ๋‚ด๋†“๋Š” ๋‹ต๋ณ€์˜ ์ผ๊ด€์„ฑ์„ ์กฐ์ ˆํ•˜๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜ 0~2๊นŒ์ง€ ์„ค์ • ๊ฐ€๋Šฅ 0์ผ ๊ฒฝ์šฐ ๋‹ต๋ณ€์ด ์ผ๊ด€์ , 2์ผ ๊ฒฝ์šฐ randomํ•˜๊ฒŒ ๋‚˜์˜ค๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Œ callbacks์˜ streaming_stdout ๋‚ด์—์„œ streaming..