๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

Statistics/Probabilistic Graphical Model

(4)
๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network)๋ฅผ ํ™œ์šฉํ•œ King County์˜ ์ง‘๊ฐ’ ๋ถ„์„ ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network)๋ฅผ ํ™œ์šฉํ•˜์—ฌ King County์˜ ์ง‘๊ฐ’์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๋‹ค์–‘ํ•œ ์š”์†Œ์˜ ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ํ™•์ธํ•˜๊ณ , ๋ถ„์„ํ•˜๋Š” ๊ฒƒ์ด ์ด ํ”„๋กœ์ ํŠธ์˜ ๋ชฉ์ ์ด๋‹ค. ๋ฐ์ดํ„ฐ์˜ ์ถœ์ฒ˜๋Š” ์—ฌ๊ธฐ(www.kaggle.com/harlfoxem/housesalesprediction)์ด๊ณ , ์•„๋ž˜์˜ ๊ตฌ์„ฑ ์ˆœ์„œ๋Œ€๋กœ ๋ถ„์„ ๋ฐ ๋ชจ๋ธ๋ง ๊ณผ์ •์„ ์ˆ˜ํ–‰ํ•  ๊ฒƒ์ด๋‹ค. ๋ชจ๋“  ์ฝ”๋“œ๋Š” R๋กœ ์ž‘์„ฑ๋˜์—ˆ๋‹ค. 1. ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ 2. ์‹œ๊ฐํ™” ๋ฐ ์ƒ๊ด€๊ด€๊ณ„ ๋ถ„์„ 3. ๋‹ค์ค‘ํšŒ๊ท€๋ถ„์„ 4. ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ ๋ชจ๋ธ๋ง 1. ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ In: # Statistic library(car) # Data manipulation library(dplyr) library(tidyr) # Visualization library(ggplot2) librar..
๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network) (3) ๋ณธ ํฌ์ŠคํŒ…์€ ์นด์ด์ŠคํŠธ ๋ฌธ์ผ์ฒ  ๊ต์ˆ˜๋‹˜์˜ ์ธ๊ณต์ง€๋Šฅ ๋ฐ ๊ธฐ๊ณ„ํ•™์Šต ๊ฐœ๋ก  2์˜ ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network) ๊ฐ•์˜ ๋‚ด์šฉ์„ ์ •๋ฆฌํ•œ ๊ฒƒ์ด๋‹ค. ๋‹ค๋ฃฐ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 1. Potential functions 2. Absorption in clique graph 3. Example of belief propagation 1. Potential functions โ–ท Potential function์€ ์ž ์žฌ์ ์œผ๋กœ ํ™•๋ฅ ์ด ๋˜๋Š” ํ•จ์ˆ˜๋กœ์จ, ์•„์ง ํ™•๋ฅ ์˜ ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์€ ํ•จ์ˆ˜๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ์ด ํ•จ์ˆ˜๋Š” Belief propagation์„ ํ†ตํ•ด ํ™•๋ฅ ๋กœ์จ ๋ฐ”๋€Œ๊ฒŒ ๋œ๋‹ค. ์ด์— ๋Œ€ํ•œ ๋‚ด์šฉ์€ ์ดํ›„์— ๋‹ค๋ฃจ๊ธฐ๋กœ ํ•˜๊ฒ ๋‹ค. โ–ท ์œ„์˜ ์˜ˆ๋Š” Potential function์„ ์„ค๋ช…ํ•˜๊ธฐ ์‰ฝ๊ฒŒ ๋‚˜ํƒ€๋‚ธ ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ..
๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network) (2) ๋ณธ ํฌ์ŠคํŒ…์€ ์นด์ด์ŠคํŠธ ๋ฌธ์ผ์ฒ  ๊ต์ˆ˜๋‹˜์˜ ์ธ๊ณต์ง€๋Šฅ ๋ฐ ๊ธฐ๊ณ„ํ•™์Šต ๊ฐœ๋ก  2์˜ ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network) ๊ฐ•์˜ ๋‚ด์šฉ์„ ์ •๋ฆฌํ•œ ๊ฒƒ์ด๋‹ค. ๋‹ค๋ฃฐ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 1. Factorization of Bayesian network 2. Conditional probability 3. Most probable assignment 4. Marginalization and elimination 5. Variable elimination 1. Factorization of Bayesian network โ–ท ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ์˜ Factorization์€ Full joint distribution์„ ๊ตฌํ•  ๋•Œ, ๊ฐœ๋ณ„ ๋…ธ๋“œ์˜ Conditional probability์˜ Condition์— ํฌํ•จ๋˜๋Š” ๋…ธ๋“œ๋ฅผ ๊ฐ..
๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network) (1) ๋ณธ ํฌ์ŠคํŒ…์€ ์นด์ด์ŠคํŠธ ๋ฌธ์ผ์ฒ  ๊ต์ˆ˜๋‹˜์˜ ์ธ๊ณต์ง€๋Šฅ ๋ฐ ๊ธฐ๊ณ„ํ•™์Šต ๊ฐœ๋ก  2์˜ ๋ฒ ์ด์ง€์•ˆ ๋„คํŠธ์›Œํฌ(Bayesian network) ๊ฐ•์˜ ๋‚ด์šฉ์„ ์ •๋ฆฌํ•œ ๊ฒƒ์ด๋‹ค. ๋‹ค๋ฃฐ ๋‚ด์šฉ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 1. Conditional vs. Marginal independence 2. Bayesian network 3. Interpretation of Bayesian network 4. Typical local structures 5. Bayes ball algorithm 1. Conditional vs. Marginal independence โ–ท Conditional independence์™€ Marginal independence๋Š” ๋‘˜ ๋‹ค ๋…๋ฆฝ์„ ์˜๋ฏธํ•˜์ง€๋งŒ, ์•ฝ๊ฐ„์˜ ์ฐจ์ด๊ฐ€ ์žˆ๋‹ค. Marginal independence๋Š” P(A|B) ..