profile-img
The merit of an action lies in finishing it to the end.
slide-image

Ensemble Learning์ด๋ž€?

์—ฌ๋Ÿฌ ๊ฐœ์˜ base-learner(base-model)๋ฅผ ์กฐํ•ฉํ•˜๋Š” ๋ชจ๋ธ

 

Bagging

- ๋‹ค๋ฅธ ์ด๋ฆ„: Bootstrap Aggregating

- Bootstrap: Random Sampling ๋ฐฉ๋ฒ•๋ก  ์ค‘ ํ•˜๋‚˜. ๋ณต์› ์ถ”์ถœ์„ ์‹œํ–‰ํ•œ๋‹ค.

- ๋ณต์› ์ถ”์ถœ์„ ํ†ตํ•ด ๋™์ผํ•œ ํฌ๊ธฐ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์—ฌ๋Ÿฌ ๊ฐœ ์ƒ์„ฑํ•œ ํ›„, ์•™์ƒ๋ธ”์„ ๊ตฌ์„ฑํ•˜๋Š” ํ•™์Šต ๋ฐฉ์‹

- ๋‹จ์ : ๋žœ๋ค ์ƒ˜ํ”Œ๋ง์€ ์šด์— ์˜์กดํ•œ๋‹ค.

๋งŒ์•ฝ, decision tree๋ฅผ ํ•™์Šต์‹œํ‚จ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด์ž.

boostrap์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”์ถœํ•  ๊ฒฝ์šฐ, ๋ณต์› ์ถ”์ถœ์ด๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ™์€ ๋ฐ์ดํ„ฐ๊ฐ€ ์—ฌ๋Ÿฌ ๋ฒˆ ๋ฝ‘ํž ์ˆ˜ ์žˆ๊ณ , ๊ทธ๋ ‡๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ์…‹๋ผ๋ฆฌ ์œ ์‚ฌํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ด๋Ÿฐ ๊ฒฝ์šฐ decision tree์˜ root node๋Š” ๋ชจ๋ธ ๊ฐ„์— ํ•ญ์ƒ ๋น„์Šทํ•ด์งˆ ์ˆ˜๋ฐ–์— ์—†๋‹ค. 

 

Boosting

- Bagging๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์ด randomness์— ์˜์กดํ•œ๋‹ค๋Š” ๋‹จ์ ์„ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋œ ๋ฐฉ์‹

- ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์„ 3๊ฐœ๋กœ ๋ถ„ํ• ํ•œ๋‹ค. (D1, D2, D3)

- ํ•™์Šต ๋ฐฉ์‹

base-learner๋ฅผ 3๊ฐœ ํ•™์Šต์‹œํ‚จ๋‹ค.

1. h1์„ D1์œผ๋กœ ํ•™์Šต์‹œํ‚จ๋‹ค.

2. h2๋ฅผ D2 ์ค‘ h1์ด ํ‹€๋ ธ๋˜ ๋ฌธ์ œ + ๋งž์•˜๋˜ ๋ฌธ์ œ ๊ฐ™์€ ๋น„์œจ๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํ•™์Šต์‹œํ‚จ๋‹ค.

3. h3์„ h2๊ฐ€ ํ‹€๋ ธ๋˜ ๋ฌธ์ œ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ํ•™์Šต์‹œํ‚จ๋‹ค.

- ์˜ˆ์ธก ๋ฐฉ์‹

h1(x)=h2(x)๋ผ๋ฉด, h1(x)๋ฅผ ๋ฐ˜ํ™˜.

์•„๋‹ˆ๋ผ๋ฉด h3(x)์„ ๋ฐ˜ํ™˜. => ํ‹€๋ ธ๋˜ ๋ฌธ์ œ๋งŒ์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์ด๋ฏ€๋กœ ๊ฐ€์žฅ ๋งž์„ ํ™•๋ฅ ์ด ๋†’๋‹ค.

1