Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
keisu_special_lecture_20210511.pdf
Search
Taro Takaguchi
May 10, 2021
Technology
0
480
keisu_special_lecture_20210511.pdf
Taro Takaguchi
May 10, 2021
Tweet
Share
More Decks by Taro Takaguchi
See All by Taro Takaguchi
takaguchi_15th_neteco.pdf
ttakaguchi
0
460
KDD2018 ダイジェスト @ Data Ship Update Lecture #6
ttakaguchi
1
100
ウェブ企業の非研究者ポジションで行うサイエンス
ttakaguchi
0
3.4k
Other Decks in Technology
See All in Technology
Intuneお役立ちツールのご紹介
sukank
3
740
AWS パートナー企業でテクニカルサポートに従事して 3年経ったので思うところをまとめてみた
kazzpapa3
1
220
Terraform Stacks入門 #HashiTalks
msato
0
220
freeeのモバイルエンジニアについて
freee
1
100
第23回Ques_タイミーにおけるQAチームの在り方 / QA Team in Timee
takeyaqa
0
190
AIチャットボット開発への生成AI活用
ryomrt
0
130
製造現場のデジタル化における課題とPLC Data to Cloudによる新しいアプローチ
hamadakoji
0
200
いろんなものと両立する Kaggleの向き合い方
go5paopao
2
970
Deno+JSRでパッケージを作って公開する
askua
0
120
3次元点群データ「VIRTUAL SHIZUOKA』のオープンデータ化による恩恵と協働の未来/FOSS4G Japan 2024
kazz24s
0
130
ドメイン名の終活について - JPAAWG 7th -
mikit
29
17k
株式会社ログラス − エンジニア向け会社説明資料 / Loglass Comapany Deck for Engineer
loglass2019
3
28k
Featured
See All Featured
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
7
560
Building a Scalable Design System with Sketch
lauravandoore
459
33k
Why You Should Never Use an ORM
jnunemaker
PRO
54
9k
Optimizing for Happiness
mojombo
376
69k
Building Applications with DynamoDB
mza
90
6.1k
Embracing the Ebb and Flow
colly
84
4.5k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
27
2k
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
159
15k
KATA
mclloyd
29
14k
Practical Orchestrator
shlominoach
186
10k
Documentation Writing (for coders)
carmenintech
65
4.4k
Building a Modern Day E-commerce SEO Strategy
aleyda
38
6.9k
Transcript
͋Δࣄۀձࣾʹ͓͚Δ σʔλαΠΤϯεͷ࣮ ߴޱ ଠ࿕ LINEגࣜձࣾ Data Science ηϯλʔ ܭֶಛผߨٛ ౦ژେֶ
ֶ෦ ܭֶՊ 2021/05/11 1
ߴޱ ଠ࿕ʢ͔͙ͨͪ ͨΖ͏ʣ LINEגࣜձࣾ Data Science ηϯλʔ γχΞσʔλαΠΤϯςΟετ / Ϛωʔδϟʔ
~2013ɹ౦ژେֶେֶӃ ใཧֶܥݚڀՊ ཧใֶઐ߈ ɹɹɹ ത࢜՝ఔʢཧใୈ̐ݚڀࣨʣ ~2017ɹࠃཱݚڀػؔʹͯϙευΫݚڀһɹ ܦྺ ࣌ͷઐ 2 ωοτϫʔΫՊֶʢಛʹ࣌ؒతʹมԽ͢ΔωοτϫʔΫʣ
اۀʹస͖͔͚ͨͬ͡ 3 2 2 2 1 1 1 3 3
4 4 3 4 ʹཱͪͦ͏ɺͦΕͰ࣮ࣾձͱͷڑԕ͍… @ LINE DEVELOPER DAY 2019 σʔλΛ׆༻ͨ͠ࣄۀͷ࠷લઢΛ ݟ͍ͨɾؔΘΓ͍ͨ
ςʔϚ ʮ͋Δࣄۀձࣾʹॴଐ͢ΔσʔλαΠΤϯςΟετ͕ɺ ͲΜͳࣄͰษڧݚڀͷܦݧΛ׆͔͍ͯ͠Δ͔ʁʯ 4
શମͷߏ 1. ରΛΔɿ ࣄۀձࣾͷσʔλαΠΤϯςΟετͬͯͲΜͳࣄʁ ʢٳܜʣ 2. தΛΔɿ ࣮ͰΑ͘༻͍Δ౷ܭͷҰ෦ͱ۩ମతͳࣄྫ 5
1. ରΛΔɿ ࣄۀձࣾͷσʔλαΠΤϯςΟετͬͯ ͲΜͳࣄʁ 6
ͦͦσʔλαΠΤϯςΟετͱʁ 7 - اۀɾ࣌ظɾίϛϡχςΟʹΑΓఆ༷ٛʑ - ಉ͡৬໊ͰҟͳΔۀɺҟͳΔ৬໊Ͱڞ௨͢Δۀ ࢦඪΛఆٛ͠ܭଌ͢Δ / ετʔϦʔΛޠΔ /
πʔϧΛ࡞Δ Analyticsʢੳܕʣ ػցֶशͷख๏ΛɾαʔϏεʹ࣮͢Δ AlgorithmsʢΞϧΰϦζϜܕʣ ౷ܭख๏ʹΑΓҼՌؔΛཱূ͢Δ Inferenceʢਪܕʣ Ref. https://www.linkedin.com/pulse/one-data-science-job-doesnt-fit-all-elena-grewal/ Data Scientist ྨͷҰྫɿ
λεΫ͝ͱʹׂ͕͔ΕΔ 8 Q1. ɾαʔϏεʹ࣮͞ΕΔίʔυΛॻ͘ʁ Analytics ʢੳܕʣ Algorithms ʢΞϧΰϦζϜܕʣ Inference ʢਪܕʣ
Q2. ౷ܭख๏ʹΑΓҼՌؔΛݕূ͢Δʁ Yes Yes No No ྨͷҰྫɿ
৫ߏ୲ྖҬʹରԠ͍ͯ͠Δ 9 Data Science ηϯλʔ Data Science Machine Learning Machine
Learning Research Analyticsʢੳܕʣ AlgorithmsʢΞϧΰϦζϜܕʣ Inferenceʢਪܕʣ جૅݚڀ͓ΑͼࣄۀͷԠ༻ ػցֶशΤϯδχΞ
ੳɾਪܕͷ۩ମతͳࣄ ਐߦத ࣄલ ࣄޙ ࣌ظ 10 Ωϟϯϖʔϯ / ৽ػೳͷՃ /
طଘػೳͷมߋͳͲ
ੳɾਪܕͷ۩ମతͳࣄ ਐߦத ࣄલ ࣄޙ ࣌ظ 11 Ωϟϯϖʔϯ / ৽ػೳͷՃ /
طଘػೳͷมߋͳͲ - Ωϟϯϖʔϯͷ݅બఆ - ৽ػೳͷχʔζݟੵΓ - ػೳมߋͷӨڹͷݟੵΓ - etc.
ੳɾਪܕͷ۩ମతͳࣄ ਐߦத ࣄલ ࣄޙ ࣌ظ 12 Ωϟϯϖʔϯ / ৽ػೳͷՃ /
طଘػೳͷมߋͳͲ - ΦϯϥΠϯ A/B ςετ - μογϡϘʔυͷ࡞ ओཁͳࣄۀࢦඪͷϞχλϦϯάද - ҟৗͳมԽͷݕग़ - etc.
ੳɾਪܕͷ۩ମతͳࣄ ਐߦத ࣄલ ࣄޙ ࣌ظ 13 Ωϟϯϖʔϯ / ৽ػೳͷՃ /
طଘػೳͷมߋͳͲ - ࢪࡦͷޮՌݕূ - ҼՌਪ - ظతมԽͷཁҼղ - etc.
νʔϜɺϓϩδΣΫτɺϓϩμΫτ 14 νʔϜ
νʔϜɺϓϩδΣΫτɺϓϩμΫτ 15 νʔϜ ϓϩδΣΫτ
νʔϜɺϓϩδΣΫτɺϓϩμΫτ 16 νʔϜ ϓϩδΣΫτ ϓϩμΫτ ྫɿϓϩμΫτʮLINE ΞϓϦʯͷ ◦◦ػೳՃϓϩδΣΫτʹؔΘΔ Data Science
νʔϜ
ʢνʔϜ|ϓϩδΣΫτ|ϓϩμΫτʣϚωδϝϯτ 17 νʔϜ ϓϩδΣΫτ ϓϩμΫτ ৫ͷඪΛઃఆ͠ɺͦͷ࣮ݱͷͨΊʹ ࿑ྗɾ࣌ؒɾ͓ۚͷΛௐ͠ޮԽ͢Δ ※ આ໌ͷͨΊʹ୯७Խ͍ͯ͠·͢
ʮయܕతͳ̍ͷࣄ༰ʁʯ 18 ࣌ظ ϓϩδΣΫτ A ϓϩδΣΫτ B ϓϩδΣΫτ C ͱ͋Δ
1 λεΫ͕ؒΛۭ͚ͯஅଓతʹਐߦ͢Δ e.g. ଞνʔϜͷਐߦͪɺಥൃతͳґཔ
ੳɾਪܕͷλεΫɿ՝ղܾͷαΠΫϧ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 19 ࣌ظ
εςοϓ̍. ؍ଌ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 20 ࣌ظ ࠷ۙɺΞΫςΟϒϢʔβʔ
͕ఀ͍ͯ͠Δʁ 2݄ 3݄ 4݄ 5݄ μογϡϘʔυɿ ओཁͳࣄۀࢦඪͷϞχλϦϯάද ※ ΓऔΓͱͯ͢Սۭͷͷ
εςοϓ̎. Ծઆͱ՝ͷઃఆ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 21 ࣌ظ ͜ͷΞΫςΟϒϢʔβʔͷ
ਪҠରॲ͖͢ͷ͔ʁ - ྫͷقઅతͳมಈʁ - Ϣʔβʔͷηάϝϯτ͝ͱͷมԽʁ - ৽ن / طଘ / ෮ؼ - ଞػೳͷར༻ϢʔβʔͷਪҠʁ → ʮ৽نϢʔβʔͷܧଓ͕Լ ͍ͯ͠Δɻݩͷਫ४ʹճ෮͢Δͱ ˓ສਓ૿ՃͷӨڹ͕͋Δʯ ※ ΓऔΓͱͯ͢Սۭͷͷ
εςοϓ̏. ղܾࡦͷཱҊ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 22 ࣌ظ -
৽نϢʔβʔʹϩάΠϯΛ ଅ͢௨ΛૹΖ͏ - ௨ͷසΛςετ͍ͨ͠ ςετͷઃܭΛ͠·͢ - ൱ΛධՁ͢Δࢦඪͷܾఆ - ςετʹඞཁͳαϯϓϧ αΠζͷܭࢉ - ൱ͷஅج४ͷ߹ҙ ※ ΓऔΓͱͯ͢Սۭͷͷ
εςοϓ̐. ݕূ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 23 ࣌ظ ςετͷ݁ՌΛੳ͠·͢
- σʔλͷਖ਼ৗͳऩूͷ֬ೝ - ࢦඪʹର͢ΔԾઆݕఆ - ՃͷվળҊͷࣔࠦ - ૯߹తͳϨϙʔςΟϯά Ճೖཌʹ̍ճ͚ͩ௨Λ ૹΔҊΛ࠾༻͢Δ ※ ΓऔΓͱͯ͢Սۭͷͷ
εςοϓ̍(2). ؍ଌ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 24 ࣌ظ ৽نϢʔβʔͷܧଓ
ࠓޙϞχλϦϯά͠·͢ ※ ΓऔΓͱͯ͢Սۭͷͷ 2݄ 3݄ 4݄ 5݄ 6݄ 2݄ 3݄ 4݄ 5݄ 6݄ μογϡϘʔυʹ߲ΛՃ͢Δ ΞΫςΟϒϢʔβʔ ৽نϢʔβʔܧଓ
ੳɾਪܕͷλεΫɿ՝ղܾͷαΠΫϧ ؍ଌ Ծઆͱ՝ ͷઃఆ ݕূ ղܾࡦͷཱҊ 25 ࣌ظ - ੳɾਪͷλεΫ
ؔऀͱͷίϛϡχέʔγϣϯΛ ௨ͯ͡ਐߦ͢Δ - ౷ܭͳͲઐࣝͷ׆༻ɺ શମͷαΠΫϧͷதͷҰཁૉ - ࠷ऴతͳҙࢥܾఆऀɺࣄۀɾ ϓϩμΫτɾϓϩδΣΫτͷऀ
ઐతͳֶͷࣝ͏ʁ 26 “LIFE AND MATHS”, © Pearls of Raw Nerdism
http://pearlsofrawnerdism.com/life-and-maths/
ઐతͳֶͷࣝ͏ʁ 27 “LIFE AND MATHS”, © Pearls of Raw Nerdism
http://pearlsofrawnerdism.com/life-and-maths/ ࢲͷߟ͑ɿ - ઐతͳֶͳ͠ͰࡁΉػձͷ΄͏͕ଟ͍ - ઐ͕ࣝ͋Δͱɺ՝ղܾͷ֤εςοϓͷ্࣭͕͕Δ
ઐతͳֶͳ͠ͰࡁΉػձͷ΄͏͕ଟ͍ ൃੜස ֶతͳ ෳࡶ 28 ߴ ߴ ֓೦ਤ
ੳɾਪܕͷׂ ˚ෳࡶͳ͜ͱΛߦ͢Δ͜ͱ ˚ཧతʹ৽نͳ͜ͱΛߦ͏͜ͱ ✓ ࣄۀʹཱͭݟΛదʹఏڙ͢Δ͜ͱ ʮࣄۀʹର͢Δߩݙʯ ʮ࣮ࢪʹཁ͢Δίετʯͷ͕࣠ӅΕ͍ͯΔ ֶతͳ͠͞ ≠ ࣄۀ্ͷ՝ղܾͷ͠͞
ฏқͳ࡞ۀɺઐతͳۀͷ ྫɿ୯७ͳूܭ࡞ۀ ઐࣝΛ ൃش͢Δۀ 29 ฏқͳ࡞ۀΛ௨ͨ͡σʔλɾࣄۀͷशख़ → ઐࣝΛൃش͢Δۀͷߦ ࣄۀʹର͢Δཧղͷ্ →
࣮ࢪʹίετͷ͔͔ΔੳλεΫͷड
ઐ͕ࣝ͋Δͱɺ՝ղܾͷ֤εςοϓͷ্࣭͕͕Δ 30 ؍ଌ Ծઆͱ՝ͷઃఆ ݕূ ղܾࡦͷཱҊ ΑΓϊΠζʹؤ݈Ͱղऍͷ͍͢͠ࢦඪΛ ఆٛͰ͖Δ ʮσʔλͱֶʹΑͬͯղ͚Δʯͷ ఆࣜԽͷϨύʔτϦʔ͕૿͑Δ
ద͔ͭޮతͳղܾࡦΛબͰ͖Δ ҙࢥܾఆʹඞཁͳݟΛత֬ʹநग़Ͱ͖Δ
֓೦ͷ֫ಘੈքͷݟ͑ํΛม͑Δ 31 ՝ɿ̎Λ̍ສݸͨ͑͠ΛΓ͍ͨ ࢉͷ֓೦ΛΒͳ͍ͱ 2 + 2 + 2 +
2 + …… ʮݱ࣮తͳ࣌ؒͰղܾͰ͖·ͤΜʯ ࢉΛ͍ͬͯΕ 2 × 10,000 = 20,000 ղ͚ͳ͍ ղ͚Δ
ʮ͑Λग़͢ͱࣄۀʹཱͭʯྖҬΛࢦ͢ 32 ࣄۀՁʹ ݁ͼͭ͘ ࣄۀՁʹ ݁ͼ͖ͭͮΒ͍ ͑Λग़ͤΔ ͑Λग़ͤͳ͍ ઐࣝͷशಘ ࣄۀͷཧղ
ؔऀͱͷର Cf. ҆ਓ, ʮΠγϡʔ͔Β͡ΊΑʕతੜ࢈ͷʰγϯϓϧͳຊ࣭ʱʯ, ӳ࣏ग़൛ʢ2010ʣ σʔλαΠΤϯςΟετͷۀ্ͷλεΫΛ̎࣍ݩʹϚοϓ͢Δ
ࣄۀձࣾͷσʔλαΠΤϯςΟετͷࣄ 33 ࣄۀͷͨΊͷ՝ղܾͷαΠΫϧ ੳܕ / ΞϧΰϦζϜܕ / ਪܕ νʔϜͱͯ͠ϓϩδΣΫτʹऔΓΉ ྨʢҰྫʣ
Ґஔ͚ͮ ੳɾਪͷλεΫ ઐతͳࣝ ՝ղܾͷ࣭Λ্͛Δ
ͲΜͳڥͩͱྗΛൃش͍͔͢͠ʁ 34 A. αΠΤϯε͕Ͱ͖Δ͜ͱ Պֶతํ๏ʹج͍ͮͯۀΛߦ͠ɺՌ͕ೝ͞ΕΔ͜ͱ - ٬؍తͳࠜڌʹج͍ͮͯɺཧΛల։͢Δ͜ͱ - खଓ͖͕ه͞Εɺ࠶ݱՄೳͰ͋Δ͜ͱ -
ͱ͘ʹ͕݁ޡΓͩͬͨ߹ʹɺݕূՄೳͰ͋Δ͜ͱ
ۀΛαΠΤϯεʹ͢ΔͨΊʹ 1. ܧଓ͢Δ 2. ԾఆΛڞ༗͢Δ 3. ࣈΛݟΔલʹஅج४ΛܾΊΔ 35 σʔλαΠΤϯςΟετଆʹ৺͕͚Δ͖͜ͱ͕͋Δ
1. ܧଓ͢Δ 36 Պֶతํ๏ɺ܁Γฦ͢͜ͱʹҙ͕ٛ͋Δ ԿΛ͖͔͢ʁ - ࠶ݱɾݕূՄೳͳΑ͏ʹهΛ͢ - ҡ࣋Մೳͳ؍ଌํ๏Λߏங͢Δ ʢϞχλϦϯάͷࣗಈԽʣ
- ࣍ͷ՝ઃఆΛଅࣔࠦ͢Λఏڙ͢Δ
2. ԾఆΛڞ༗͢Δ 37 ܦݧՊֶʹ͓͚ΔՊֶతࣝ ✗ ઈରෆมͷਅ࣮ͷू߹ ✓ ؍ଌͱԾఆʹج͍ͮͯਪ͞Εͨؼ݁ ԿΛ͖͔͢ʁ -
ԾఆΛ໌֬ʹ͑Δ ʮϢʔβʔͷ૿Ճઌ݄ͱಉ͡ͱԾఆ͠·͢ʯ - ݕূͷεςοϓͰɺࣄલͷԾఆͷଥੑݕূ͢Δ ʮϢʔβʔͷ૿Ճɺ݁Ռతʹઌ݄ͱൺͯʙͰͨ͠ʯ
3. ࣈΛݟΔલʹஅج४ΛܾΊΔ 38 ྔ → ৗݴޠͷมʹᐆດੑ͕͋Δ ͜ͷࢦඪ͕ “ेʹ” ্ঢͨ͠Β ςετޭͱஅ͠·͠ΐ͏
ʢ+3% ”े” ͩΖ͏͔…ʣ ԿΛ͖͔͢ʁ - ࣄલʹஅج४ΛྔతʹܾΊΔ - ج४ͷࠜڌ٬؍తʹ͢Δ (ྫ) ࣄۀඪʹର͢Δظد༩ ɹɹ͡ΒΕͨίετͷճऩ ɹɹաڈͷྨࣅࣄྫͷ݁Ռ ※ ͯ͢Սۭͷͷ ࢦඪͷ্ঢ +3% Ͱͨ͠
ٳܜ 39
શମͷߏ 1. ରΛΔɿ ࣄۀձࣾͷσʔλαΠΤϯςΟετͬͯͲΜͳࣄʁ ʢٳܜʣ 2. தΛΔɿ ࣮ͰΑ͘༻͍Δ౷ܭͷҰ෦ͱ۩ମతͳࣄྫ 40
2. தΛΔɿ ࣮ͰΑ͘༻͍Δ౷ܭͷҰ෦ͱ۩ମతͳࣄྫ 41
ੳɾਪͷ۩ମతͳࣄʢ࠶ܝʣ 42 ਐߦத ࣄલ ࣄޙ ࣌ظ Ωϟϯϖʔϯ / ৽ػೳͷՃ /
طଘػೳͷมߋͳͲ ΦϯϥΠϯ A/B ςετ 1. αϯϓϧαΠζͷܭࢉ 2. ଟॏൺֱ 3. ׳ΕޮՌͷਪఆ
1. αϯϓϧαΠζͷܭࢉ 43 ςετରͷࠩΛݕఆ͢ΔͨΊʹඞཁͳαϯϓϧαΠζΛࢉग़͢Δ͜ͱ Q. αϯϓϧαΠζΛܭࢉܾͯ͠ΊΔཧ༝ʁ A. ౷తͳԠ༻Ͱɺαϯϓϧऩूͷίετ͕ߴ͔ͬͨ ɹe.g. ྟচݚڀɺۀ
Q. ΣϒαʔϏεͳΒαϯϓϧऩूͷίετߴ͘ͳ͍ͷͰʁ
ΣϒαʔϏεͰαϯϓϧαΠζΛܭࢉ͢Δཧ༝ 44 1. աʹେ͖ͳαϯϓϧαΠζ → খ͞ͳมԽͰ༗ҙʹͳΓ͕ͪ ʮ౷ܭతʹ༗ҙʯڧ͍ҹΛ༩͑Δදݱ 2. ಛʹςετҊ͕ྑ͘ͳ͍࣌ɺϢʔβʔʹແ༻ͳӨڹΛ༩͑ͯ͠·͏ 4.
P-Hacking ͷ༨͕Δ ʮ༗ҙ͕ࠩग़ͳ͔͔ͬͨΒαϯϓϧαΠζΛେ͖ͯ͘͠࠶ςετ͠Α͏ʯ 3. SUTVA (Stable Unit Treatment Value Assumption) ͕ഁΕ͘͢ͳΔ ʮ͋ΔϢʔβʔͷߦಈଞͷϢʔβʔͷׂΓͯʹӨڹ͞Εͳ͍ʯ ʢྫʣςετը໘͕ڞ༗͞ΕΔɺϝσΟΞʹऔΓ্͛ΒΕΔ
αϯϓϧαΠζܭࢉͷجຊܗ 45 ઃఆ - ಠཱͳ̎܈αϯϓϧͷฏۉͷݕఆ - ࢄ̎܈Ͱಉ͡ & ط -
αϯϓϧαΠζઍ ~ ສ݅ఔऔΕΔ ݕఆ͞ΕΔԾઆ - ؼແԾઆ - ରཱԾઆ H0 H1 μ1 − μ2 = 0 μ1 − μ2 ≠ 0 αϯϓϧαΠζɹͷܾఆʹඞཁͳύϥϝʔλ - ༗ҙਫ४ - ݕग़ྗ - ޮՌྔ - ࢄ α 1 − β δ = μ1 − μ2 σ2 < + ∞ n ʢɹ ͕ਅͷ߹ʣ H1
αϯϓϧαΠζܭࢉͷ෮शʢ̍ʣ 46 ਤԼهจݙΑΓ࠶ߏͨ͠ Gerald van Belle, “Statistical Rules of Thumb”
(2nd edition), Wiley, 2008 ඪຊฏۉͷࠩ x1 − x2 H0 : μ1 − μ2 = 0 0 S . E . = σ 2 n ਖ਼نͷ࠶ੜੑΑΓ α 2 α 2 ༗ҙਫ४ɹɿ α ɹ͕ਅͷͱ͖ɹ Λ࠾ͯ͠͠·͏֬ ʢِཅੑʣ H0 H1
αϯϓϧαΠζܭࢉͷ෮शʢ̎ʣ 47 ਤԼهจݙΑΓ࠶ߏͨ͠ Gerald van Belle, “Statistical Rules of Thumb”
(2nd edition), Wiley, 2008 ඪຊฏۉͷࠩ x1 − x2 H1 : μ1 − μ2 = δ δ S . E . = σ 2 n H0 : μ1 − μ2 = 0 0 β = 1− ݕग़ྗ (1 − β) ɹ͕ਅͷͱ͖ɹ Λ࠾ͯ͠ ͠·͏֬ʢِӄੑʣ H0 H1 β
αϯϓϧαΠζܭࢉͷ෮शʢ̏ʣ 48 ਤԼهจݙΑΓ࠶ߏͨ͠ Gerald van Belle, “Statistical Rules of Thumb”
(2nd edition), Wiley, 2008 ඪຊฏۉͷࠩ x1 − x2 δ S . E . = σ 2 n 0 β n* = 2σ2 (z1−α/2 + z1−β) 2 δ2 㱺 ཁ݅Λຬͨͨ͢Ίʹ ࠷ݶඞཁͳαϯϓϧαΠζ z1−α/2 σ 2 n* = δ − z1−β σ 2 n* ඪ४ਖ਼نͷ Ґؔ α 2
1. ਅͷͷࢄɹ ͕େ 2. ِཅੑɺِӄੑΛ͑Δ ɹɹ͕খ 3. ݕग़͍ͨ͠ޮՌྔɹ͕খ ͕େ͖͘ͳΔཁҼ n*
σ2 α, β δ αϯϓϧαΠζʹ͍ͭͯͷิ 49 n* = 2σ2 (z1−α/2 + z1−β) 2 δ2 ཁ݅Λຬͨͨ͢Ίʹ ࠷ݶඞཁͳαϯϓϧαΠζ ύϥϝʔλͷܾΊํʢҰྫʣ ɹɹɿ׳शతͳ͔ɺݫ͠ʹ ɹɹɿۙͷ࣮ଌ ɹɹʢςετޙʹଥੑΛ֬ೝʣ ɹɹɿ׳शతͳ ɹɹɹor ίετΛ্ճΔޮՌ ɹɹɹor աڈͷྨࣅ͢Δςετ݁Ռ α, β σ2 δ
2. ଟॏൺֱ 50 ࣮Ͱɺ̏܈Ҏ্ͷൺֱΛٻΊΒΕΔ͜ͱ͕Α͋͘Δ 㲗 લઅͰ෮शͨ̎͠܈ؒͷݕఆ എܠ ظؒͰͳΔ͘ଟ͘ͷՄೳੑΛςετ͍ͨ͠ ܈ؒϖΞͷճ͚ͩ୯७ʹݕఆΛ܁Γฦ͍͚ͯ͠ͳ͍ ʂʂ
ଟॏൺֱ ͷ →
࣮ྫɿάϧʔϓ࡞ը໘ͷมߋςετ 51 “ίϛϡχέʔγϣϯΞϓϦʮLINEʯͷػೳվળΛࢧ͑ΔσʔλαΠΤϯε” LINE DEVELOPER DAY 2019 https://linedevday.linecorp.com/jp/2019/sessions/B1-3 άϧʔϓ࡞ͷखॱΛɺΑΓ͍͘͢վྑ͍ͨ͠
̎ͭͷมߋΛΈ߹Θͤͯࢼ͍ͨ͠ 52 “ίϛϡχέʔγϣϯΞϓϦʮLINEʯͷػೳվળΛࢧ͑ΔσʔλαΠΤϯε” LINE DEVELOPER DAY 2019 https://linedevday.linecorp.com/jp/2019/sessions/B1-3 1. ࠷ۙτʔΫͨ͠༑ͩͪΛ༏ઌදࣔ͢Δ
2. खॱΛ̍ը໘ʹ·ͱΊΔ
̎×̎=̐௨Γ̒ϖΞͷݕఆʁ 53 “ίϛϡχέʔγϣϯΞϓϦʮLINEʯͷػೳվળΛࢧ͑ΔσʔλαΠΤϯε” LINE DEVELOPER DAY 2019 https://linedevday.linecorp.com/jp/2019/sessions/B1-3 গͷީิͰɺৄࡉͳݕূҙ֎ͳ΄ͲෳࡶʹͳΔ
ݕఆͷ܁Γฦ͠Կ͕͔ʁ 54 ݕఆ͞ΕΔԾઆ - ؼແԾઆ - ରཱԾઆ H0 H1 θ1
= θ2 = θ3 = θ4 ʢ̐܈ͷ߹ʣ {θi} i=1,2,3,4 ͷ͏ͪগͳ͘ͱ̍ͭͷϖΞͰ θi ≠ θj (i ≠ j) ࣮ߦతͳ༗ҙਫ४ Family-Wise Error Rate α = 1 − (1 − α)6 ≥ α α α 1 − (1 − α)6 શମͱͯ͠ݟͨ࣌ʹɺِཅੑ্͕͕ͬͯ͠·͏
Bonferroni ิਖ਼ 55 ֤ϖΞͷݕఆͷ༗ҙਫ४ΛɺݕఆͷճɹͰׂͬͨʹௐ͢Δ α → α m m Family-Wise
Error Rate α ≤ α ͱͳΓɺશମͱͯ͠ͷ༗ҙਫ४͕อͨΕΔ σϝϦοτ ͕େ͖͍ͱอकతʹͳΓ͕ͪʢِӄੑͷ্ঢʣ m
άϧʔϓ࡞ը໘ςετͰͷରॲ 56 “ίϛϡχέʔγϣϯΞϓϦʮLINEʯͷػೳվળΛࢧ͑ΔσʔλαΠΤϯε” LINE DEVELOPER DAY 2019 https://linedevday.linecorp.com/jp/2019/sessions/B1-3 - ࣄલݕূʹج͍ͮͯɺରΛ̏܈̎ϖΞʹߜΔ
- ̎ϖΞʹରͯ͠ Bonferroni ิਖ਼ͯ͠ݕఆ͢Δ α → α/2
άϧʔϓ࡞ը໘ςετͷ݁Ռ 57 “ίϛϡχέʔγϣϯΞϓϦʮLINEʯͷػೳվળΛࢧ͑ΔσʔλαΠΤϯε” LINE DEVELOPER DAY 2019 https://linedevday.linecorp.com/jp/2019/sessions/B1-3 ʮ̎ը໘ +
࠷ۙτʔΫͨ͠༑ͩͪϦετʯ → ࡞ྃΛҡ࣋ͭͭ͠ɺ࡞ͷॴཁ࣌ؒΛॖͨ͠
ଟॏൺֱͷରॲʹऔΓೖΕ͍ͯΔ͜ͱ 58 - جຊతʹൺֱରΛߴʑ̐ύλʔϯʹݶఆ͢Δ ̑ύλʔϯҎ্ੳղऍ͘͠ͳΔ - Bonferroni ิਖ਼ͰِཅੑΛ੍͢Δ ϦεΫͷ͋ΔςετͰِཅੑΛආ͚͍ͨ -
σϝϦοτΛิ͏ͨΊɺݕग़ྗɹΛߴΊʹઃఆ͢Δ β
3. ׳ΕޮՌͷਪఆ 59 ΦϯϥΠϯςετͷظؒ௨ৗ̎ʙ̏िؒ ظؒͷԠΛͦͷ··ड͚औͬͯΑ͍ͷͩΖ͏͔ʁ Ծઆ ಛʹྺ࢙͕͘श׳Խ͍ͯ͠Δػೳ΄Ͳɺ ը໘ͷมߋʹର͢ΔҰ࣌తͳԠ͕ݱΕΔ ՝ Ұ࣌తͳԠ͕ఆৗతͳར༻ʹམͪண͘
ʮ׳ΕޮՌʯΛݕग़͍ͨ͠
࣮ྫɿ༑ͩͪՃը໘ͷγϯϓϧԽςετ 60 “ίϛϡχέʔγϣϯΞϓϦʮLINEʯʹ͓͚Δ࣮ફతσʔλαΠΤϯε” DEIM 2020 https://engineering.linecorp.com/ja/blog/deim2020-report/ - ༑ͩͪՃը໘͔Βɺ༑ͩͪՃҎ֎ͷΞΠςϜΛআ͢Δ - ༑ͩͪՃ
& LINEެࣜΞΧϯτՃ͕ݮগͯ͠͠·ͬͨ
LINE ͷ৽نϢʔβʔͷΈʹݶఆͯ͠ܭଌͯ͠ΈΔ 61 - ༑ͩͪՃɾLINEެࣜΞΧϯτՃͱʹ༗ҙͳมԽͳ͠ - ԾઆɿશମʹطଘϢʔβʔ͕׳ΕΔ·ͰͷԠ͕ݱΕ͍ͯΔʁ “ίϛϡχέʔγϣϯΞϓϦʮLINEʯʹ͓͚Δ࣮ફతσʔλαΠΤϯε” DEIM 2020
https://engineering.linecorp.com/ja/blog/deim2020-report/
׳ΕޮՌΛࠩͷࠩͰϞσϧԽ͢Δ 62 1st half 2nd half Control Treatment yT,1 yC,1
yC,2 yT,2 ׳ΕޮՌҎ֎ͷӨڹ̎܈ͰಉҰ ʢฒߦτϨϯυ & ڞ௨γϣοΫͷԾఆʣ Ծఆ ςετظؒΛલɾޙʹ̎͢Δ ࠩͷࠩ౷ܭྔ δ = (yT,2 − yC,2) − (yT,1 − yC,1) ճؼϞσϧԽ ̂ β3 = ̂ δ y = β0 + β1 T + β2 S + β3 TS + ε T / C ͷμϛʔ 1st / 2nd ͷμϛʔ Ͱ͋Γɺ ճؼϞσϧͷͯ·Γ & ͷ༗ҙੑΛ֬ೝ͢Δ
༑ͩͪՃը໘ͷγϯϓϧԽςετͷ݁Ռ 63 - LINEެࣜΞΧϯτՃͷมԽʹɺ׳ΕޮՌ͕ݱΕ͍ͯͨ - LINEެࣜΞΧϯτՃͷͷΈআͯ͠ɺϦϦʔε͞Εͨ “ίϛϡχέʔγϣϯΞϓϦʮLINEʯʹ͓͚Δ࣮ફతσʔλαΠΤϯε” DEIM 2020 https://engineering.linecorp.com/ja/blog/deim2020-report/
ख๏Λඪ४Խͯ͠ਫฏల։͢Δ 64 “σʔλαΠΤϯε͕ಋ͘τʔΫϝχϡʔUIͷϦχϡʔΞϧϓϩδΣΫτ” LINE DEVELOPER DAY 2020 https://linedevday.linecorp.com/2020/ja/sessions/3932 - ׳ΕޮՌݕग़๏ɺτʔΫϝχϡʔͷϦχϡʔΞϧͰ׆༻ͨ͠
- ۀΛαΠΤϯεʹ͢ΔͨΊʹ - ܧଓ͢Δ
࣮ͰΑ͘༻͍Δ౷ܭͷҰ෦ͱ۩ମతͳࣄྫ 65 ΦϯϥΠϯ A/B ςετ 1. αϯϓϧαΠζͷܭࢉ 2. ଟॏൺֱ 3.
׳ΕޮՌͷਪఆ
͓ΘΓʹ 66
ʮσʔλαΠΤϯςΟετʯͷকདྷʁ 67 ࢲͷߟ͑ɿ - ʮσʔλαΠΤϯςΟετʯͱ͍͏ݺশظมΘ͍͔ͬͯ͘ - σʔλͱֶΛͬͯ՝Λղܾ͢Δͱ͍͏ཁ໘ന͞ ˠ ݺশͷਰΑΓͣͬͱ͘ଓͩ͘Ζ͏
ੈք͔ΒࣝΛநग़͢ΔαΠΫϧ 68 ੈք σʔλ ࣝ ॲཧ݁Ռ ࣝΛͲ͏ੈքʹϑΟʔυόοΫ͢Δ͔ʁ ੈքʹͲ͏͋ͬͯ΄͍͔͠ͱ͍͏Ձஅ ԿΛɺԿͷͨΊʹͲ͏؍ଌ͢Δ͔ʁ ܭଌʹ͢Δ͔൱͔ͱ͍͏Ձஅ
ରͷੈքΛݶఆ͢ΕࣗಈԽͷՄೳੑ͕͋Δ 69 ੈք σʔλ ࣝ ॲཧ݁Ռ ͷΓग़͠ͱγεςϜԽɺιϑτΣΞΤϯδχΞϦϯάͷྖҬ ʢྫʣهࣄͷਪનɺϚʔέςΟϯά ՁஅͷॏཁੑΓଓ͚Δ
ֶੜͷօ͞Μͷϝοηʔδ 70 ֶͼଓ͚·͠ΐ͏ औΓΈ·͠ΐ͏ - ֶ෦ɾେֶӃͰͷݚڀʢ՝ղܾͷαΠΫϧʣ ɹେֶੈքϨϕϧͷઐՈ͔ΒֶΔوॏͳॴ - ֶɾֶͷઐࣝ -
ϓϩάϥϛϯά - ޠֶ - ٕज़ྙཧɺ๏੍ɺྺ࢙ɺ… ৬໊τϐοΫͷྲྀߦʹͱΒΘΕ͗ͣ͢ɺ ઐࣝͰ՝ղܾͰ͖ΔਓΛͥͻࢦ͍ͯͩ͘͠͞