$30 off During Our Annual Pro Sale. View Details »

機械学習とのつきあいかた / How to get involved with Machine Learning @ Wantedly

機械学習とのつきあいかた / How to get involved with Machine Learning @ Wantedly

2019年新人研修で使った資料です。

以下の内容を話した全体的にはポエムです。

1. Wantedlyの機械学習
2. MLと組織
3. MLとデータ
4. MLプロジェクの進め方

Makoto Tanji

April 24, 2019
Tweet

More Decks by Makoto Tanji

Other Decks in Programming

Transcript

  1. ػցֶशͱͷ͖͍͔ͭ͋ͨ
    @Wantedly
    New Grad Training 2019
    April 24, 2019 - Makoto Tanji (@tan-z-tan)

    View Slide

  2. ©2019 Wantedly, Inc.
    • ML Engineer at Wantedly, Inc.
    • Wantedly Visit 2015 - 2016
    • Client Growth, Scout
    • Wantedly People 2016 - current
    • Server side
    • Machine Learning
    • Data Analysis
    ࣗݾ঺հ Who am I?
    Makoto Tanji

    View Slide

  3. ©2019 Wantedly, Inc.
    ࠓ೔࿩͢͜ͱ

    View Slide

  4. ©2019 Wantedly, Inc.
    1. ςʔϚ
    2. Wantedlyͷػցֶश
    3. Take home messages
    1. MLͱ૊৫
    2. MLͱσʔλ
    3. MLϓϩδΣΫͷਐΊํ
    4. ·ͱΊ
    Contents

    View Slide

  5. ©2019 Wantedly, Inc.
    ϝσΟΞʹࡌΔΑ͏ͳ࠷৽ٕज़ͷػցֶशͷ׆༻͸՚΍͔Ͱ͕͢ɺ࣮ӡ༻
    ͷͨΊʹ͸࣮ࡍ͸๲େͳσʔλ͕ඞཁͩͬͨΓɺਫ਼౓͕ग़ͳ͔ͬͨΓɺ҆
    ఆͨ͠ӡ༻͕೉͍͠ͱݴ͏ࠔ೉΋ଘࡏ͠·͢ɻ
    8BOUFEMZͷ.-ΤϯδχΞ͸ɺ8ΤϯδχΞɾσʔλαΠΤϯςΟε
    τɾΠϯϑϥͱڠಇ͠ͳ͕Β͜ͷ໰୊ʹऔΓ૊ΜͰ͍·͢ɻ
    ·ͨকདྷ8ΤϯδχΞ΋ΠϯϑϥΤϯδχΞ΋ػցֶशͱԿΒ͔ͷܗͰ
    ؔΘΔ͜ͱ͕ࠓΑΓ΋૿͖͑ͯ·͢ɻ
    8BOUFEMZͰ࣮ࡍʹಈ͍͍ͯΔػցֶशΛݟͳ͕Βػցֶशͱͷ͖͍ͭ͋
    ͔ͨΛߟ͑ͯߦ͖·͠ΐ͏ɻ
    ςʔϚ

    View Slide

  6. ©2019 Wantedly, Inc.
    શମతʹػցֶशʹ·ͭΘΔϙΤϜͰ͢

    View Slide

  7. ©2019 Wantedly, Inc.
    Wantedlyͷػցֶश׆༻ࣄྫ
    Machine Learnings working at Wantedly

    View Slide

  8. ©2019 Wantedly, Inc.
    Wantedlyͷػցֶश׆༻ࣄྫ

    View Slide

  9. ©2019 Wantedly, Inc.
    Wantedlyͷػցֶश׆༻ࣄྫ
    7JTJU
    w ืूͷਪન
    w 1VTI௨஌ͷύʔιφϥΠζ
    w εΧ΢τϑΟϧλͷվળ
    w ʜ
    1FPQMF
    w ໊ࢗը૾ͷ0$3ͱ෼ྨ
    w ਓͷਪન
    w هࣄͷਪનɺ1VTIͷ։෧཰༧ଌ
    w ༷ʑͳݻ༗දݱநग़
    w ʢػցֶशͷϦϙδτϦ͸.-λά͕͍ͭͯ·͢ʣ
    w IUUQTHJUIVCDPNXBOUFEMZ VUG&$RNMUZQFMBOHVBHF

    View Slide

  10. ©2019 Wantedly, Inc.
    ͳͥػցֶशΛಋೖ͢Δͷ͔ʁ

    View Slide

  11. ©2019 Wantedly, Inc.
    Wantedlyͷػցֶश׆༻ࣄྫ
    ఏڙ͍ͨ͠Ձ஋͕͋Δ͔Β
    • εΩϟϯ͔Β਺ඵͰσʔλԽͱ͍͏ମݧΛఏڙ͍ͨ͠
    w ໊ࢗΛຕຕ࿈བྷઌʹҠ͢ͷਏ͍ˠͦΕਓؒͷ࢓ࣄ͡Όͳ͍Ͱ͢ɻΧϝϥͰࡱͬͨΒҰ౓ʹσʔλԽ
    • ίίϩΦυϧγΰτʹग़ձ͏ମݧΛఏڙ͍ͨ͠
    w ͳΜ͔͍͍࢓ࣄͳ͍͔ͳʁˠ͋ͳ͕ͨڵຯ͕͋Γͦ͏ͳืू͔ΒࣗಈͰਪન͠·͢Α

    View Slide

  12. ©2019 Wantedly, Inc.
    ࣾ಺Ͱಈ͍͍ͯΔ.-ͷߏ੒ྫ

    View Slide

  13. ©2019 Wantedly, Inc.
    ࣾ಺Ͱಈ͍͍ͯΔMLͷߏ੒ྫ
    ֶशࡁΈͷ"1*ύλʔϯ
    ਪનύλʔϯ
    ͦͷଞ

    View Slide

  14. ©2019 Wantedly, Inc.
    ֶशࡁΈͷAPIύλʔϯ
    ϚΠΫϩαʔϏεͷαʔόͷ"1*ͱͯ͠ఏڙ
    1. OCRʢจࣈೝࣝʣAPIɺςΩετ͔Βͷநग़APIɺಉҰੑ൑ఆAPIͳͲ୔ࢁ
    2. લ΋ֶͬͯशͨ͠ϞσϧΛಈ͔ͯ݁͠ՌΛฦ͢
    3. e.g. Peopleͷ໊ࢗೝࣝؔ࿈API͸͜ͷྫ͕ଟ͍
    ߟ͑Δ͜ͱ
    • ਫ਼౓͕஌Βͳ͍͏ͪʹམ͍ͪͯͳ͍͔
    • ϝτϦΫεΛຖ೔औΔɻ৽͘͠APIΛ࡞Δͱ͖͸νΣοΫϦετʹೖΕΔ
    • ίετ
    • CPU཯଎ʹͳ͍ͬͯΔ͜ͱ͕ଟ͘ɺkubernetesͷϦιʔεΛେྔʹ࢖ͬͯͳ͍͔

    View Slide

  15. ©2019 Wantedly, Inc.
    ਪનύλʔϯ
    લ΋͓ͬͯ͢͢ΊείΞΛܭࢉɾఏڙ
    1. Ϣʔβ͕ϖʔδʹΞΫηεͨ͠ͱ͖ʹҰॠͰΦεεϝͷืूΛग़͍ͨ͠
    2. Ϣʔβͷաڈͷߦಈ͔Βࣄલʹܭࢉ͓ͯ͘͠
    3. Ϣʔβ×ืूͷ݁ՌΛσʔλετΞʹอଘ
    ߟ͑Δ͜ͱ
    • ʮϢʔβ×ืूʯ͕͸͍ΔσʔλετΞ
    • ຖ೔ͷֶशJob͕ࢭ·͍ͬͯͳ͍͔ʁ
    • Job͕ଟஈʹͳΔͱಛʹΘ͔Γʹ͘͘ͳΔˠData Pipelineͷ࿩

    View Slide

  16. ©2019 Wantedly, Inc.
    ͦͷଞ
    σʔλͷਖ਼نԽ
    1. ࡶଟͳ৘ใΛਖ਼نԽɾಉҰੑͳͲΛ൑ఆ͢Δ
    2. ࣙॻΛ࡞Δ࢓ࣄ
    3. Ϣʔβʹ͸ݟ͑ͳ͍͚Ͳ಺෦Ͱݡ͘ͳ͍ͬͯΔ
    #JH2VFSZ.-
    1. ಺෦ͷϩδοΫͰ࢖͍ͬͯΔ
    2. ࣮͸࢖ΘΕͯ·͢

    View Slide

  17. ©2019 Wantedly, Inc.
    ࠓ೔ͷTake-Home Messages

    View Slide

  18. ©2019 Wantedly, Inc.
    1. ૊৫ͷ࿩
    2. σʔλͷ࿩
    3. ϓϩδΣΫτͷਐΊํͷ࿩

    View Slide

  19. ©2019 Wantedly, Inc.
    ૊৫ͷ࿩

    View Slide

  20. ©2019 Wantedly, Inc.
    ʮ૊৫ߏ଄͸ઓུΛܾΊΔʯ ΠΰʔϧɾΞϯκϑ

    ʢ͔͍͍ͬ֨͜ݴΛݴ͍͔͚ͨͬͨͩʣ

    View Slide

  21. ©2019 Wantedly, Inc.
    ૊৫ͷ࿩
    8BOUFEMZͰ͸ɺϓϩμΫτνʔϜʹػցֶश
    ΤϯδχΞ͕͍Δ
    MLΤϯδχΞ
    ̏ਓ
    MLΤϯδχΞ
    ̎ਓ
    Infrastructure
    Visit People

    View Slide

  22. ©2019 Wantedly, Inc.
    ૊৫ͷ࿩
    Ϣʔβ·Ͱͷڑ཭͕͍ۙ
    1. ߦͬͨվળ͕μΠϨΫτʹϢʔβ·Ͱಧ͘
    ϓϩμΫτʹՁ஋ͷ͋ΔվળΛߦ͍΍͍͢
    1. ϓϩμΫτͷ޲͔͍ͬͯΔํ޲Λڞ༗͍ͯ͠ΔͷͰᴥᴪ͕ى͜ΓͮΒ͍
    2. ඞཁͰ͋Ε͹ϓϩμΫτͷϏδωεϩδοΫ·ͰखΛೖΕΒΕΔ
    վળ͚ͩͰͳ͘৽͍͠ػೳΛࣗવͱٞ࿦͢Δ
    1. ։ൃνʔϜͷҰһͳͷͰ਺஋తͳվળ͚ͩͰͳ͘ʮ͜͏͋Δ΂͖ʯํ޲Λٞ࿦ͯ͠ਐΊΒΕΔ

    View Slide

  23. ©2019 Wantedly, Inc.
    ૊৫ͷ࿩
    ଞͷ૊৫ͷ͋Γํ
    σʔλαΠΤϯςΟετ
    ΤϯδχΞ
    σβΠφ
    Ϛωʔδϟ
    શࣾԣஅ
    MLνʔϜ
    ϓϩμΫτνʔϜ
    શࣾԣஅ
    σʔλ෼ੳνʔϜ
    MLΤϯδχΞ
    ґཔ
    ݁Ռ

    View Slide

  24. ©2019 Wantedly, Inc.
    ૊৫ͷ࿩
    ଞͷ։ൃνʔϜͱ͸Ͳ͏ؔΘΔͷ͔ʁ
    1. ։ൃελΠϧ: APIΛఆٛͯ͠WEBଆͱMLଆͰฒߦͯ͠։ൃ
    1. e.g. Ϣʔβʹෳ਺ͷهࣄΛฦ͍ͨ͠ɻهࣄͷϦετ͸WEBଆͰऔಘ͢ΔɻMLͷAPIΛݺͼग़ͯ͠ॱংΛܾఆ͢Δ
    2. ؔΘΓํ
    1. WEBଆ: લ΋ͬͯAPI spec͚ܾͩΊ͓ͯ͘ͱɺSTUBͯ͠MLଆΛ଴ͨͣʹ։ൃՄೳɻ
    2. Πϯϑϥ: service-in͢ΔલʹɺෛՙͷϝτϦΫεͱ࠷ѱࢮΜͰ΋͍͍API͔ΫϦςΟΧϧͳAPI͔Λ͢Γ߹Θͤ
    3. ίϛϡχέʔγϣϯ
    1. ࣄલʹʮਫ਼౓ͷظ଴஋ʯΛ͢Γ߹ΘͤΔͱ͍͍ɻ
    2. ௒ݡ͍API͕αΫοͱͰ͖Δ͜ͱ͸΄΅ແ͍ɻ99.9ˋͷਫ਼౓͕ඞཁͳཁ݅ͳͷ͔ɺݟͤํΛڞ༗͢Δ

    View Slide

  25. ©2019 Wantedly, Inc.
    8BOUFUEMZͰ͸ɺϓϩμΫτνʔϜʹ.-ΤϯδχΞ͕ೖΔͱ͍͏ಇ͖ํΛ
    औ͍ͬͯ·͢ɻϢʔβʹۙ͘ɺϓϩμΫτυϦϒϯͳಇ͖ํ͕͠΍͍͢ͱ
    ͍͏ϝϦοτ͕͋Γ·͢ɻ
    ϓϩμΫτνʔϜͷதͰ͸ɺࣄલʹʮޓ͍ͷ੹೚ൣғʯͱʮਫ਼౓ͷظ଴஋ʯ
    Λ͸͖ͬΓͤ͞Δͱࠞཚͳ͘։ൃ͕ਐΈ·͢ɻ
    ૊৫ͷ࿩

    View Slide

  26. ©2019 Wantedly, Inc.
    %FFQ-FBSOJOHྠߨձ΍ͬͯ·͢
    ຖिਫ༵೔ʙ!8BOUFEMZ

    View Slide

  27. ©2019 Wantedly, Inc.
    σʔλ

    View Slide

  28. ©2019 Wantedly, Inc.
    σʔλ
    ໘ന͍ྫ͑ʮσʔλ͸ݪ༉ʯ
    ͦͷ··Ͱ͸࢖͑ͳ͍ɻਫ਼੡͕ඞཁ
    Ճ޻͢Δͱ༷ʑͳ༻్ʹ࢖͑Δ
    ੈքதͰσʔλͷऔΓ߹͍Λ͍ͯ͠Δ
    ग़లɿෆ໌

    View Slide

  29. ©2019 Wantedly, Inc.
    Wantedlyͷ໘ന͍σʔλ
    w ϢʔβͷϓϩϑΟʔϧ਺ඦສΦʔμʔ
    w Ϣʔβͷͭͳ͕ΓωοτϫʔΫΤοδ਺ԯΦʔμʔ
    w ಡΈऔΒΕ໊ͨࢗԯ͘Β͍

    View Slide

  30. ©2019 Wantedly, Inc.
    Data lake
    ඞཁͳσʔλ͸#JH2VFSZʹ
    ू໿Ͱ͖Δ࢓૊Έ͕͋Δ
    "OBMZUJDT

    View Slide

  31. ©2019 Wantedly, Inc.
    ؾΛ͚ͭΔ͜ͱ
    w σʔλपΓͰ࣮ࡍ͋ͬͨ໰୊
    1. RPAͷ׆༻Ͱҙຯͷͳ͍େྔͷΞΫηε͕͋ͬͨͨΊɺݟ͔͚ͷฏۉར༻਺্͕͕ͬ
    ͨɻίϯόʔδϣϯΛ൐Θͳ͍ϩά͕૿͑ͯ਺஋͕มԽͨ͠
    2. ৽͍͠ػೳͷϦϦʔεͰςʔϒϧͷΧϥϜʹ৽͍͠஋͕ೖΔΑ͏ʹͳͬͨɻલ΋ͬͯ
    ݕ஌ͨͨ͠ΊϦϦʔεલʹमਖ਼Ͱ͖͕ͨɺ஌Βͳ͔ͬͨΒϞσϧ͕յΕ͍͔ͯͨ΋

    View Slide

  32. ©2019 Wantedly, Inc.
    ؾΛ͚ͭΔ͜ͱ
    w σʔλͷϓϥΠόγʔͷ໰୊
    1. ίϯϓϥΠΞϯεΛकΓ·͠ΐ͏

    View Slide

  33. ©2019 Wantedly, Inc.
    8BOUFEMZͰ͸େن໛ͳϢʔβؔ࿈ͷσʔλΛ#JH2VFSZʹू໿͞ΕΔͱ͍
    ͏ஈ֊·Ͱ࣮ݱ͞Ε͓ͯΓɺσʔλΛ࢖͏͜ͱࣗମ͸༰қʹͳΓ·ͨ͠ɻ
    ࠓޙ͸ߋʹਐΜͰ׆༻͠΍͕͢͞ΛղܾʹͳΓ·͢ɻ
    σʔλ΍ϩά͸αʔόͷ࣮૷͚ͩʹ׬݁ͯ͠ͳ͘ɺϝτϦΫε΍σʔλΛ
    ࢖͏.-ͷϞσϧʹ΋ӨڹΛ༩͑·͢ɻαʔόͷมߋ͕.-ϞσϧΛյ͞ͳ
    ͍Α͏ʹपΓʹڞ༗͠·͠ΐ͏ɻʢຊ౰͸ҟৗݕ஌ͳͲͰγεςϜతʹक
    Γ͍ͨʣ
    σʔλ

    View Slide

  34. ©2019 Wantedly, Inc.
    ϓϩδΣΫτͷਐΊํ

    View Slide

  35. ©2019 Wantedly, Inc.
    ػցֶशͷϓϩδΣΫτ͸Ͳ͏΍ͬͯ࢝·Δͷʁ

    View Slide

  36. ©2019 Wantedly, Inc.
    έʔεόΠέʔεͰҰൠԽ͕೉͍͠
    খ͍͞ϓϩδΣΫτ͸ࡶஊϕʔε͕
    ࢝·Γͷ͜ͱ͕ଟ͍ؾ͕͢Δ

    View Slide

  37. ©2019 Wantedly, Inc.
    ੈͷதͷ#FTU1SBDUJDFΛ୳ͦ͏

    View Slide

  38. ©2019 Wantedly, Inc.
    Machine Learning Best Practice
    • ϧʔϧ1: ػցֶशΛ࢖Θͳ͍Ͱ࢝ΊΒΕͳ͍͔ߟ͑Δ
    • ϧʔϧ2: ઃܭͯ͠ࢦඪΛ࡞ΔʢݱࡏͷγεςϜΛཧղ͔ͯ͠ΒࢦඪΛ࡞Δʣ
    • ϧʔϧ3: ϩδοΫ͕ෳࡶʹͳͬͬͨΒػցֶशΛબ୒͢Δ
    ref: https://developers.google.com/machine-learning/guides/rules-of-ml/

    View Slide

  39. ©2019 Wantedly, Inc.
    ϓϩδΣΫτͷਐΊํ
    ྫɿʮ͓஌Γ߹͍Ͱ͔͢ʁʯػೳ

    View Slide

  40. ©2019 Wantedly, Inc.
    ϓϩδΣΫτͷਐΊํ
    ख୳Γظ ϧʔϧ૿Ճظ ػցֶश
    • ڞ௨ͷ஌ਓ਺͕ଟ͍ਓʢΛਪન͢Δʣ
    • ϝτϦΫε=ϦΫΤετ਺ͱঝೝ཰
    • ڞ௨ͷ஌ਓ਺͕ଟ͍ਓ
    • ಉ͡ձࣾͷਓ
    • ڞ௨ͷ໊ࢗΛ͍࣋ͬͯΔ਺͕ଟ͍ਓ
    • ಉ࣌ʹεΩϟϯ͞ΕͨਓͰڞ௨ͷͭ
    ͳ͕Γ͕͋Δ
    • ͜ͷ૊Έ߹Θͤ
    • ɾɾɾ
    • Deep Learningϕʔεͷਪન
    • + ABςετ

    View Slide

  41. ©2019 Wantedly, Inc.
    ϓϩδΣΫτͷਐΊํ
    ػցֶशΛಋೖ͢Δͱ͖ʹ͸
    े෼ෳࡶͰݱঢ়ͷࢦඪΛվળͰ͖ͦ͏ͳ՝୊ΛબͿ

    View Slide

  42. ©2019 Wantedly, Inc.
    ·ͱΊ

    View Slide

  43. ©2019 Wantedly, Inc.
    ·ͱΊ
    ػցֶशͷϓϩμΫτ͸ɺ8ΤϯδχΞɾΠ
    ϯϑϥνʔϜͱҰॹʹ࡞͍ͬͯΔ
    ϓϩμΫτνʔϜͷதʹ.-ΤϯδχΞ͕͍Δߏ
    ଄ͳͷͰɺҰॹʹٞ࿦ͯ͠։ൃͰ͖Δ؀ڥ
    ద੾ͳ՝୊Λݟ͚ͭͯ.-ͰՁ஋Λग़͢͜ͱʹڵ
    ຯΛ޿͛ͯ΄͍͠

    View Slide

  44. ©2019 Wantedly, Inc.
    References
    • Rules of Machine Learning: Best Practices for ML Engineering
    • https://developers.google.com/machine-learning/guides/rules-of-ml/
    • Wantedly ͷػցֶशϓϩμΫτ։ൃΛࢧ͑Δػցֶशج൫ / #rejectcon2018
    • https://speakerdeck.com/south37/number-rejectcon2018
    Photo Credit
    • https://unsplash.com/photos/PMwu9gfCSbw

    View Slide