๐Ÿ’ก TIL

    Uber๊ฐ€ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋„์ฐฉ์‹œ๊ฐ„์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ๋ฒ•

    Uber๊ฐ€ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋„์ฐฉ์‹œ๊ฐ„์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ๋ฒ•

    DeepETA: How Uber Predicts Arrival Times Using Deep Learning Uber๊ฐ€ ๋„์ฐฉ์‹œ๊ฐ„์„ ์˜ˆ์ธกํ•˜๊ธฐ ์œ„ํ•ด ์™œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ๋กœ ๋„˜์–ด๊ฐ€๊ฒŒ ๋˜์—ˆ๋Š”์ง€ ์•Œ ์ˆ˜ ์žˆ๋Š” ๊ธ€์ด๋‹ค. ์•„๋ž˜ ๋ณธ๋ฌธ์€ ์ด๋ฅผ ์ฝ๊ณ  ์ •๋ฆฌํ•œ ๋‚ด์šฉ์ด๋‹ค. ์šฐ๋ฒ„๋Š” ์‚ฌ์šฉ์ž์—๊ฒŒ ์ฐจ๋Ÿ‰ ๋„์ฐฉ ์˜ˆ์ธก ์‹œ๊ฐ„(ETA)์„ ์ œ๊ณตํ•œ๋‹ค. ETA๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์š”๊ธˆ์„ ๊ณ„์‚ฐํ•˜๊ณ , ํ”ฝ์—… ์‹œ๊ฐ„์„ ์ถ”์ •ํ•˜๊ณ , ๋ผ์ด๋”์™€ ๊ธฐ์‚ฌ๋ฅผ ์—ฐ๊ฒฐํ•˜๊ณ , ๋ฐฐ์†ก์„ ๊ณ„ํšํ•˜๋Š” ๋“ฑ์˜ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ์ด๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์ถ”์ •ํ•˜๋Š” ๋Šฅ๋ ฅ์€ ๋งค์šฐ ์ค‘์š”ํ•˜๋‹ค. ์ •ํ™•ํ•œ ETA ์ถ”์ •์€ ๊ณ ๊ฐ์—๊ฒŒ ๊ธ์ •์ ์ธ ๊ฒฝํ—˜์„ ์ œ๊ณตํ•˜๊ณ , ์„œ๋น„์Šค ๊ฐ€๊ฒฉ๊ณผ ์šด์ „ ๊ฒฝ๋กœ ๋“ฑ์„ ์„ค์ •ํ•˜๋Š” ๋ฐ๋„ ํ™œ์šฉ๋œ๋‹ค. ์ด ๊ธ€์—์„  ์šฐ๋ฒ„๊ฐ€ ETA ์˜ˆ์ธก ๊ฐœ์„ ์„ ์œ„ํ•ด ์™œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์„ ํƒํ–ˆ๊ณ , ์–ด๋–ค ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ–ˆ๋Š”์ง€ ๋‹ค๋ฃฌ๋‹ค. ์ „ํ†ต์ ์ธ ETA ์—”์ง„..

    Decision Tree

    Decision Tree

    ์–ด๋ ต๊ณ  ๋ณต์žกํ•œ ๋ชจ๋ธ์ด ์•„๋‹Œ ๊ธฐ๋ณธ ํŠธ๋ฆฌ๋ชจ๋ธ์ด ์„ฑ๋Šฅ์ด ๋” ์ข‹์€ ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค. ๋˜ ์ธ๊ธฐ์žˆ๋Š” ๋ถ€์ŠคํŒ… ๋ชจ๋ธ์ธ Light GBM, XG boost์€ Tree๋ชจ๋ธ์„ ๊ณ ๋„ํ™”ํ•œ ๊ฒƒ์ด๊ณ , ๋‚˜ ๋˜ํ•œ ํ”„๋กœ์ ํŠธ์—์„œ ์‚ฌ์šฉํ–ˆ๋˜ ๋ชจ๋ธ์ด๊ธฐ์— ๋‹ค์‹œ ํ•œ ๋ฒˆ ์ •ํ™•ํžˆ ์ดํ•ดํ•ด๋ณด๊ณ ์ž StatQuest: Decision Trees๋ฅผ ๋ณด๊ณ  ์ •๋ฆฌํ•ด๋ดค๋‹ค. * ์ด ๊ธ€์€ StatQuest: Decision Trees๋ฅผ ๋ณด๊ณ  ์ •๋ฆฌํ•œ ๋‚ด์šฉ์ž…๋‹ˆ๋‹ค. * 0. Decision Tree Decision Tree๋Š” ๋ถ„๋ฅ˜ ๊ทœ์น™์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„๋ฅ˜, ํšŒ๊ท€ํ•˜๋Š” ์ง€๋„ ํ•™์Šต ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ด๋‹ค. ํ•œ ๋ฒˆ์˜ ๋ถ„๊ธฐ ๋•Œ๋งˆ๋‹ค ๋ณ€์ˆ˜ ์˜์—ญ์„ ๋‘ ๊ฐœ๋กœ ๊ตฌ๋ถ„ํ•˜๋ฉฐ, ์งˆ๋ฌธ์ด๋‚˜ ์ •๋‹ต์€ ๋…ธ๋“œ(Node)๋ผ๊ณ  ๋ถˆ๋ฆฐ๋‹ค. ๋งจ ์ฒ˜์Œ ๋ถ„๋ฅ˜ ๊ธฐ์ค€์„ Root Node, ์ค‘๊ฐ„ ๋ถ„๋ฅ˜ ๊ธฐ์ค€์„ Intermediate N..

    Hidden Technical Debt in Machine Learning Systems (2)

    Hidden Technical Debt in Machine Learning Systems (2)

    Hidden Technical Debt in Machine Learning Systems Part of Advances in Neural Information Processing Systems 28 (NIPS 2015) Abstract Machine learning offers a fantastically powerful toolkit for building useful complexprediction systems quickly. This paper argues it is dangerous to think ofthese quick wins as coming for free. Using the software engineering frameworkof technical debt, we find it is..

    Hidden Technical Debt in Machine Learning Systems (1)

    Hidden Technical Debt in Machine Learning Systems (1)

    Hidden Technical Debt in Machine Learning Systems Part of Advances in Neural Information Processing Systems 28 (NIPS 2015) Abstract Machine learning offers a fantastically powerful toolkit for building useful complexprediction systems quickly. This paper argues it is dangerous to think ofthese quick wins as coming for free. Using the software engineering frameworkof technical debt, we find it is..

    ์–ผ๊ตด์ธ์‹์—์„œ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ๋ฅผ ์œ„ํ•œ PCA ์ ‘๊ทผ

    ์–ผ๊ตด์ธ์‹์—์„œ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ๋ฅผ ์œ„ํ•œ PCA ์ ‘๊ทผ

    PCA๊ฐ€ ์˜์ƒ์ธ์‹์— ํ™œ์šฉ๋˜๋Š” ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋Š” ์–ผ๊ตด์ธ์‹(face recognition)์ž…๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จ๋œ ๊ฐœ๋… ํ˜น์€ ์šฉ์–ด๋กœ์„œ eigenface(์•„์ด๊ฒํŽ˜์ด์Šค)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์€ 20๊ฐœ์˜ 45x40 ์–ผ๊ตด ์ด๋ฏธ์ง€๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€์—์„œ ํ”ฝ์…€ ๋ฐ๊ธฐ๊ฐ’์„ ์ผ๋ ฌ๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ ๋ฒกํ„ฐ๋กœ ๋งŒ๋“ค๋ฉด, ์ด๋“ค ๊ฐ๊ฐ์˜ ์–ผ๊ตด ์ด๋ฏธ์ง€๋Š” 45x40 = 1,800 ์ฐจ์›์˜ ๋ฒกํ„ฐ๋กœ ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (์ฆ‰, ๊ฐ๊ฐ์˜ ์ด๋ฏธ์ง€๋Š” 1,800 ์ฐจ์› ๊ณต๊ฐ„์—์„œ ํ•œ ์ (์ขŒํ‘œ)์— ๋Œ€์‘) ์ด์ œ ์ด 20๊ฐœ์˜ 1,800์ฐจ์› ์  ๋ฐ์ดํ„ฐ๋“ค์„ ๊ฐ€์ง€๊ณ  PCA๋ฅผ ์ˆ˜ํ–‰ํ•˜๋ฉด ๋ฐ์ดํ„ฐ์˜ ์ฐจ์› ์ˆ˜์™€ ๋™์ผํ•œ ๊ฐœ์ˆ˜์˜ ์ฃผ์„ฑ๋ถ„ ๋ฒกํ„ฐ๋“ค์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์–ป์–ด์ง„ ์ฃผ์„ฑ๋ถ„ ๋ฒกํ„ฐ๋“ค์„ ๋‹ค์‹œ ์ด๋ฏธ์ง€๋กœ ํ•ด์„ํ•œ ๊ฒƒ์ด eigenface ์ž…๋‹ˆ๋‹ค. (์–ผ๊ตด ์ด๋ฏธ์ง€๋ฅผ ๊ฐ€์ง€๊ณ  ์–ป์€ ๋ฒกํ„ฐ์ด๊ธฐ์— ei..