Supervised vs. unsupervised learning

์ง€๋„ํ•™์Šต vs. ๋น„์ง€๋„ํ•™์Šต

Supervised learning (์ง€๋„ํ•™์Šต)

  • ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์ฃผ์ž…ํ•˜๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋ ˆ์ด๋ธ”์ด๋ผ๋Š” ์›ํ•˜๋Š” ๋‹ต์ด ํฌํ•จ
  • ๋ถ„๋ฅ˜(Classification)๊ฐ€ ์ „ํ˜•์ ์ธ ์ง€๋„ ํ•™์Šต
  • ์˜ˆ์ธก ๋ณ€์ˆ˜(predicator variable)๋ผ ๋ถ€๋ฅด๋Š” ํŠน์„ฑ(feature)์„ ์‚ฌ์šฉํ•ด target ์ˆ˜์น˜๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ = ํšŒ๊ท€(Regression)
  • ์ผ๋ถ€ ํšŒ๊ท€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ถ„๋ฅ˜์— ์‚ฌ์šฉ, ๋˜๋Š” ์ผ๋ถ€ ๋ถ„๋ฅ˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํšŒ๊ท€์— ์‚ฌ์šฉ
    e.g ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€ : ํด๋ž˜์Šค์— ์†ํ•  ํ™•๋ฅ ์„ ์ถœ๋ ฅ

๐Ÿ”Ž ์ฃผ์š” ์ง€๋„ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜
โ—พ k-์ตœ๊ทผ์ ‘ ์ด์›ƒ (k-nearest neighbors)
โ—พ ์„ ํ˜• ํšŒ๊ท€ (linear regression)
โ—พ ๋กœ์ง€์Šคํ‹ฑ ํšŒ๊ท€ (logistic regression)
โ—พ ์„œํฌํŠธ ๋ฒกํ„ฐ ๋จธ์‹  (support vector machine)
โ—พ ๊ฒฐ์ • ํŠธ๋ฆฌ์™€ ๋žœ๋ค ํฌ๋ ˆ์ŠคํŠธ (decision tree and random forest)
โ—พ ์‹ ๊ฒฝ๋ง (neural networks)


Unsupervised learning (๋น„์ง€๋„ํ•™์Šต)

  • ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋ ˆ์ด๋ธ”์ด ์—†์Œ
  • ์‹œ์Šคํ…œ์ด ์•„๋ฌด๋Ÿฐ ๋„์›€ ์—†์ด ํ•™์Šตํ•ด์•ผ ํ•จ
  • ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์Šค์Šค๋กœ ๋ฐฉ๋ฌธ์ž ์‚ฌ์ด์˜ ์—ฐ๊ฒฐ๊ณ ๋ฆฌ๋ฅผ ์ฐพ์Œ
  • ๊ณ„์ธต ๊ตฐ์ง‘ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜๋ฉด ๊ฐ ๊ทธ๋ฃน์„ ๋” ์ž‘์€ ๊ทธ๋ฃน์œผ๋กœ ์„ธ๋ถ„ํ™”
  • ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์žˆ๋Š” ์—ฌ๋Ÿฌ ํŠน์„ฑ์„ ํ•˜๋‚˜์˜ ํŠน์„ฑ์œผ๋กœ ํ•ฉ์น  ์ˆ˜ ์žˆ์Œ

๐Ÿ”Ž ์ฃผ์š” ๋น„์ง€๋„ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜
โ—พ ๊ตฐ์ง‘ (clustering)
k-ํ‰๊ท  (k-means)
DBSCAN
๊ณ„์ธต ๊ตฐ์ง‘ ๋ถ„์„ (HCA)
์ด์ƒ์น˜ ํƒ์ง€์™€ ํŠน์ด์น˜ ํƒ์ง€
์›-ํด๋ž˜์Šค (one-class SVM)
์•„์ด์†”๋ ˆ์ด์…˜ ํฌ๋ ˆ์ŠคํŠธ
โ—พ ์‹œ๊ฐํ™”์™€ ์ฐจ์› ์ถ•์†Œ
์ฃผ์„ฑ๋ถ„ ๋ถ„์„ (PCA)
์ปค๋„ PCA
์ง€์—ญ์  ์„ ํ˜• ์ž„๋ฒ ๋”ฉ (LLE)
t-SNE
โ—พ ์—ฐ๊ด€ ๊ทœ์น™ ํ•™์Šต
์–ดํ”„๋ผ์ด์–ด๋ฆฌ
์ดํด๋ ›


โ˜บ Outlier detection(์ด์ƒ์น˜ ํƒ์ง€) vs. novelty detection(ํŠน์ด์น˜ ํƒ์ง€)
  • ์ด์ƒ์น˜ ํƒ์ง€ : ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ด์ƒํ•œ ๊ฐ’์„ ์ž๋™์œผ๋กœ ์ œ๊ฑฐ
  • ํŠน์ด์น˜ ํƒ์ง€ : ํ›ˆ๋ จ ์„ธํŠธ์— ์žˆ๋Š” ๋ชจ๋“  ์ƒ˜ํ”Œ๊ณผ ๋‹ฌ๋ผ ๋ณด์ด๋Š” ์ƒˆ๋กœ์šด ์ƒ˜ํ”Œ์„ ํƒ์ง€ํ•˜๋Š” ๊ฒƒ์ด ๋ชฉ์ . ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ๊ฐ์ง€ํ•˜๊ณ  ์‹ถ์€ ๋ชจ๋“  ์ƒ˜ํ”Œ์„ ์ œ๊ฑฐํ•œ ๋งค์šฐ ๊นจ๋—ํ•œ ํ›ˆ๋ จ ์„ธํŠธ ํ•„์š”
โ˜บ ์ฐจ์› ์ถ•์†Œ(dimensionality reduction)์™€ ํŠน์„ฑ ์ถ”์ถœ(feature extraction)
  • ์ฐจ์› ์ถ•์†Œ : ๋„ˆ๋ฌด ๋งŽ์€ ์ •๋ณด๋ฅผ ์žƒ์ง€ ์•Š์œผ๋ฉด์„œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ„์†Œํ™” –> ์—ฌ๋Ÿฌ ํŠน์„ฑ์„ ํ•˜๋‚˜๋กœ ํ•ฉ์น˜๊ธฐ = ํŠน์„ฑ ์ถ”์ถœ


Semisupervised learning (์ค€์ง€๋„ํ•™์Šต)

  • ์ผ๋ถ€๋งŒ ๋ ˆ์ด๋ธ”์ด ์žˆ๋Š” ๋ฐ์ดํ„ฐ
  • ์ง€๋„ ํ•™์Šต๊ณผ ๋น„์ง€๋„ ํ•™์Šต์˜ ์กฐํ•ฉ
    e.g ์‹ฌ์ธต ์‹ ๋ขฐ ์‹ ๊ฒฝ๋ง (DBN)์€ ์—ฌ๋Ÿฌ ๊ฒน์œผ๋กœ ์Œ“์€ ์ œํ•œ๋œ ๋ณผ์ธ ๋งŒ ๋จธ์‹  (RBM)์ด๋ผ ๋ถˆ๋ฆฌ๋Š” ๋น„์ง€๋„ ํ•™์Šต์— ๊ธฐ์ดˆ. RBM์ด ๋น„์ง€๋„ ํ•™์Šต ๋ฐฉ์‹์œผ๋กœ ์ˆœ์ฐจ์ ์œผ๋กœ ํ›ˆ๋ จ๋œ ๋‹ค์Œ ์ „์ฒด ์‹œ์Šคํ…œ์ด ์ง€๋„ ํ•™์Šต ๋ฐฉ์‹์œผ๋กœ ์„ธ๋ฐ€ํ•˜๊ฒŒ ์กฐ์ •

Reinforcement learning (๊ฐ•ํ™”ํ•™์Šต)

  • ํ•™์Šตํ•˜๋Š” ์‹œ์Šคํ…œ = agent
  • ํ™˜๊ฒฝ์„ ๊ด€์ฐฐํ•ด์„œ ํ–‰๋™์„ ์‹คํ–‰ํ•˜๊ณ  ๊ทธ ๊ฒฐ๊ณผ๋กœ ๋ณด์ƒ์„ ๋ฐ›์Œ
  • ์‹œ๊ฐ„์ด ์ง€๋‚˜๋ฉด์„œ ๊ฐ€์žฅ ํฐ ๋ณด์ƒ์„ ์–ป๊ธฐ ์œ„ํ•ด ์ •์ฑ…์ด๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ์ตœ์ƒ์˜ ์ „๋žต์„ ์Šค์Šค๋กœ ํ•™์Šต
  • ์ •์ฑ…์€ ์ฃผ์–ด์ง„ ์ƒํ™ฉ์—์„œ ์—์ด์ „ํŠธ๊ฐ€ ์–ด๋–ค ํ–‰๋™์„ ์„ ํƒํ•ด์•ผ ํ• ์ง€ ์ •์˜


โ˜… ๋” ๋งŽ์€ ๋‚ด์šฉ ๋ณด๋Ÿฌ๊ฐ€๊ธฐ click click! โ˜