๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ
TIL๐Ÿ”ฅ/๋ฉ‹์Ÿ์ด์‚ฌ์ž์ฒ˜๋Ÿผ_AI School 5๊ธฐ

[๋ฉ‹์‚ฌ] AI SCHOOL 5๊ธฐ_ Day 30

by hk713 2022. 4. 12.

Decision Tree(์˜์‚ฌ๊ฒฐ์ •๋‚˜๋ฌด)

์ถœ์ฒ˜_ https://ratsgo.github.io/machine%20learning/2017/03/26/tree/

์ดํ•ดํ•˜๊ธฐ ์‰ฝ๊ณ  ํ•ด์„๋„ ์šฉ์ดํ•˜๋‹ค. ํ•˜์ง€๋งŒ 2๊ฐ€์ง€ ํฐ ๋‹จ์ ์ด ์žˆ๋‹ค

  • ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์˜ ์ž‘์€ ๋ณ€๋™์—๋„ Tree์˜ ๊ตฌ์„ฑ์ด ํฌ๊ฒŒ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋‹ค
  • ๊ณผ์ ํ•ฉ์ด ์‰ฝ๊ฒŒ ๋ฐœ์ƒํ•œ๋‹ค

→ ๊ทธ๋ž˜์„œ ๋ชจ๋ธ ์•™์ƒ๋ธ” ๊ธฐ๋ฒ• ์ค‘ ํ•˜๋‚˜์ธ Boosting ๊ธฐ๋ฒ•์ด ์ ์šฉ๋˜์—ˆ๋‹ค.

 

AdaBoost (Adaptive Boosting)

์ถœ์ฒ˜_ https://subscription.packtpub.com/book/big_data_and_business_intelligence/9781788295758/4/ch04lvl1sec32/adaboost-classifier

1) ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์—ฌ๋Ÿฌ weak learner๋“ค์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ƒ์„ฑํ•œ๋‹ค

2) ์•ž์„  learner๊ฐ€ ์ž˜๋ชป ์˜ˆ์ธกํ•œ ๋ฐ์ดํ„ฐ์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜๋„๋ก ํ•˜๊ณ (boosting) ํ•™์Šต์‹œํ‚จ๋‹ค

3) ์ตœ์ข…์ ์œผ๋กœ ๋งŒ๋“ค์–ด์ง„ strong learner๋ฅผ ์ด์šฉํ•˜์—ฌ ์‹ค์ œ ์˜ˆ์ธก์„ ์ง„ํ–‰ํ•œ๋‹ค

์—ฌ๊ธฐ์„œ ๋ฌธ์ œ๋Š” ๋†’์€ ๊ฐ€์ค‘์น˜๋ฅผ ๊ฐ€์ง„ data point๊ฐ€ ์กด์žฌํ•˜๊ฒŒ ๋˜๋ฉด ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ๋–จ์–ด์กŒ์Œ

→ ๊ทธ๋ž˜์„œ ์–ด๋–ป๊ฒŒ ํ•˜๋ฉด ์—๋Ÿฌ๋ฅผ ์ตœ์†Œํ™”ํ•˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ๊ฐ€์ค‘์น˜๋ฅผ ๋งค๊ฒจ์ค„ ์ˆ˜ ์žˆ์„๊นŒ ๊ณ ๋ฏผํ•˜๋‹ค๊ฐ€ ๋‚˜์˜จ๊ฒŒ..

 

Gradient Boosing(2012)

๊ฒฝ์‚ฌ ํ•˜๊ฐ•๋ฒ•์„ ์‚ฌ์šฉํ•ด AdaBoost์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ–ˆ๋‹ค. 

๋ฌธ์ œ๋Š” ํ•™์Šต ์„ฑ๋Šฅ์€ ์ข‹์€๋ฐ, ๋ชจ๋ธ์˜ ํ•™์Šต ์‹œ๊ฐ„์ด ์˜ค๋ž˜ ๊ฑธ๋ ธ๋‹ค

-> ๊ทธ๋ž˜์„œ ๋˜ ๊ณ ๋ฏผํ•˜๋‹ค๊ฐ€ ๋‚˜์˜จ๊ฒŒ..

 

XG Boost (Extreme Gradient Boosting, 2016)

๋ณ‘๋ ฌ์ฒ˜๋ฆฌ๋ฅผ ํ•จ์œผ๋กœ์จ ์‹œ๊ฐ„์„ ๋‹จ์ถ•์‹œ์ผฐ๋‹ค!

๊ทธ๋ฆฌ๊ณ  ์—ฌ๊ธฐ์„œ ๋” ์—…๊ทธ๋ ˆ์ด๋“œ ๋ผ์„œ Light GBM(2018)๋„ ๋งŒ๋“ค์–ด์กŒ๋‹ค!


SVM (Support Vector Machines, ์„œํฌํŠธ ๋ฒกํ„ฐ ๋จธ์‹ )

SVM์€ Margin์„ ์ตœ๋Œ€ํ™”ํ•˜๋Š” ๊ฒฐ์ • ๊ฒฝ๊ณ„(๋ฉด)๋ฅผ ์ฐพ๋Š” ๊ธฐ๋ฒ•์ด๋‹ค.

๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ์— ๋”ฐ๋ผ ์„ ํ˜• ๋ถ„๋ฆฌ๊ฐ€ ๋ถˆ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค.

๊ทธ๋Ÿด ๋•Œ๋Š” original data๊ฐ€ ๋†“์—ฌ์žˆ๋Š” ์ฐจ์›์„ ๋น„์„ ํ˜• ๋งคํ•‘(Mapping)์„ ํ†ตํ•ด ๊ณ ์ฐจ์› ๊ณต๊ฐ„์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋‹ค!

Kernelized Support Vector Machine 

์ปค๋„ ํ•จ์ˆ˜๋„ ์ง์ ‘ ๊ณจ๋ผ์„œ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์— Hyper-parameter ๋‹ค!๊ทผ๋ฐ ์—ฌ๊ธฐ์„œ ์ปค๋„ํ•จ์ˆ˜๋„ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ปดํ“จํ„ฐ๊ฐ€ ์ข‹์€ ๊ฑธ๋กœ ์•Œ์•„์„œ ์ฐพ์•˜์œผ๋ฉด ์ข‹๊ฒ ๋Š”๋ฐ... ๋ผ๋Š” ์ƒ๊ฐ์„ ๊ตฌํ˜„ํ•œ๊ฒŒ ๋ฐ”๋กœ ๋”ฅ๋Ÿฌ๋‹์ด๋‹ค..!

๋Œ“๊ธ€