๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

l1๊ทœ์ œ #l2๊ทœ์ œ #๋”ฅ๋Ÿฌ๋‹์ •๊ทœํ™” #regularization #๊ณผ์ ํ•ฉ๋ฐฉ์ง€ #๋จธ์‹ ๋Ÿฌ๋‹ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ #ridgeํšŒ๊ท€ #lassoํšŒ๊ท€ #weightdecay #elasticnet1

L1 vs L2 ๊ทœ์ œํ™” ์ฐจ์ด ์ •๋ฆฌ! ๊ณผ์ ํ•ฉ ๋ฐฉ์ง€ ํ•ต์‹ฌ ๊ธฐ๋ฒ• ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋‹ค ๋ณด๋ฉด ๊ฐ€์žฅ ํ”ํžˆ ๋งˆ์ฃผ์น˜๋Š” ๋ฌธ์ œ๊ฐ€ ๊ณผ์ ํ•ฉ(overfitting)์ž…๋‹ˆ๋‹ค. ์ด๋•Œ ํšจ๊ณผ์ ์œผ๋กœ ๊ณผ์ ํ•ฉ์„ ์ค„์ด๋Š” ๋Œ€ํ‘œ์ ์ธ ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜๊ฐ€ ๋ฐ”๋กœ ๊ทœ์ œํ™”(Regularization)์ž…๋‹ˆ๋‹ค.ํŠนํžˆ ๋จธ์‹ ๋Ÿฌ๋‹๊ณผ ๋”ฅ๋Ÿฌ๋‹์—์„œ ๊ฐ€์žฅ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๊ทœ์ œํ™” ๊ธฐ๋ฒ•์ด L1 ๊ทœ์ œ์™€ L2 ๊ทœ์ œ์ธ๋ฐ์š”, ์ด๋ฒˆ ๊ธ€์—์„œ๋Š” ์ด ๋‘ ๊ฐ€์ง€ ๊ทœ์ œ ๋ฐฉ๋ฒ•์˜ ๊ฐœ๋…, ์ˆ˜ํ•™์  ์ •์˜, ์ฐจ์ด์ , ์–ธ์ œ ์‚ฌ์šฉํ•˜๋ฉด ์ข‹์€์ง€๋ฅผ ์ •๋ฆฌํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.๐Ÿ“š ๋ชฉ์ฐจ1. ๊ทœ์ œํ™”(Regularization)๋ž€?2. L1 ๊ทœ์ œ(Lasso)๋ž€?3. L2 ๊ทœ์ œ(Ridge)๋ž€?4. L1 vs L2 ๊ทœ์ œ์˜ ์ฐจ์ด์ 5. ์–ธ์ œ L1, ์–ธ์ œ L2๋ฅผ ์‚ฌ์šฉํ• ๊นŒ?6. ๋งˆ๋ฌด๋ฆฌ ์š”์•ฝ1. ๊ทœ์ œํ™”(Regularization)๋ž€?๊ทœ์ œํ™”๋Š” ๋ชจ๋ธ์ด ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๊ณผํ•˜๊ฒŒ ์ ํ•ฉ(overfit)๋˜๋Š” ๊ฒƒ์„ ๋ฐฉ.. 2025. 4. 15.

ํ‹ฐ์Šคํ† ๋ฆฌํˆด๋ฐ”