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
Prophetを使ったコスパの良い時系列データ予測 / prophet-use-cases
Search
Haruki Okuyama
August 28, 2019
Business
0
100
Prophetを使ったコスパの良い時系列データ予測 / prophet-use-cases
ビジネスにおける時系列データをトレンド・季節性・イベント効果の3つに分けて予測する
Haruki Okuyama
August 28, 2019
Tweet
Share
More Decks by Haruki Okuyama
See All by Haruki Okuyama
Prophetを使った時系列データ予測と機械学習モデルとの比較 / prophet-vs-ml
spring1018
0
1.2k
Other Decks in Business
See All in Business
Backlogに登録するタスクの大きさについて
kozotaira
0
110
経験は色褪せない ~ キャリアの “点” を “線” にするために ~
natty_natty254
4
160
27.02.2025 El mercado cuartohorario de electricidad
neuroenergia
PRO
0
160
Clarity for Product People
arnekittler
0
1.3k
不確実性の高い仮説を 迅速に検証するための開発プロセス
mkitahara01985
2
340
Lisse/採用ピッチ資料
lisse
0
800
ホットリンク会社紹介スライド / We are hiring
hottolink_pr
1
9k
HRBP+のご紹介
masakisukeda
0
880
GA technologies Co.,Ltd. Corporate Story
gatechnologies
2
1.1k
Leading Mark新卒採用資料
unno
0
620
株式会社リブセンス 会社説明資料(報道関係者様向け)
livesense
PRO
0
1.1k
unnameカルチャーブック 2025.02.21 update
unnameinc
6
17k
Featured
See All Featured
A Tale of Four Properties
chriscoyier
158
23k
How to Think Like a Performance Engineer
csswizardry
22
1.4k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
134
33k
The Straight Up "How To Draw Better" Workshop
denniskardys
232
140k
VelocityConf: Rendering Performance Case Studies
addyosmani
328
24k
Java REST API Framework Comparison - PWX 2021
mraible
29
8.4k
Bash Introduction
62gerente
611
210k
Side Projects
sachag
452
42k
The Language of Interfaces
destraynor
156
24k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
129
19k
A better future with KSS
kneath
238
17k
Building Applications with DynamoDB
mza
93
6.3k
Transcript
©2019 Wantedly, Inc. ProphetΛͬͨ ίεύͷྑ͍࣌ܥྻσʔλ༧ଌ Creators MeetUp ϏΞόογϡLTձ Aug 28,
2019 - Haruki Okuyama - @spring1018x
©2019 Wantedly, Inc. Self-Introduction •Haruki OkuyamaʢԞࢁ ݰكʣ •Wantedly, Inc. (since
April 2019) •Recommendation Team •Like puyopuyo-tetris
©2019 Wantedly, Inc. ͢͜ͱ ɾ࣌ܥྻσʔλಛ༗ͷੑ࣭ ɾ࣌ܥྻ༧ଌϥΠϒϥϦProphetΛͬͨ༧ଌ ͞ͳ͍͜ͱ ɾProphetͷΞϧΰϦζϜ ɾProphetͷύϥϝʔλɾػೳͷৄࡉ About
this talk
©2019 Wantedly, Inc. KPI࣌ܥྻσʔλͰ͋Δ͜ͱ͕΄ͱΜͲ ɾKPIͷඪΛઃఆ͍ͨ͠ ݱࡏͷɾ݄͝ͱͷมಈΛߟྀͯ͠༧ଌͨ͠ʹ͢Δඞཁ͕͋Δ ɾKPIͷੑ࣭Λ͔ͭΈ͍ͨ KPI͕มԽͨ͠ࡍʹ, ԿͷӨڹʹΑΔͷ͔ΛఆྔతʹѲ͍ͨ͠ ࣌ܥྻσʔλͷ༧ଌ
©2019 Wantedly, Inc. ɾτϨϯυ ɾقઅੑ ɾex)िɾ݄৳ͼΔ, 3݄ब׆γʔζϯ ɾٳɾॕޮՌ ɾΠϕϯτΩϟϯϖʔϯ ࣌ܥྻσʔλͷ༧ଌͷԿ͕͍͠ʁ
͜ΕΒ͕ෳࡶʹࠞ͟Δ
©2019 Wantedly, Inc. Wantedlyͷ࣌ܥྻσʔλͷྫ ɾμογϡϘʔυมಈ͕େ͖͍σʔλͷ߹, τϨϯυ͕Θ͔Γʹ͍͘ ɾٳͷӨڹΛड͚͍͕͢, िɾ݄ʹΑͬͯٳͷ͕ҟͳΔ ɾिɾ݄ຖͷ͕ఆྔతʹΘ͔͍ͬͯͳ͍ͱͷஅ͕Ͱ͖ͳ͍ ex)
3݄ब׆γʔζϯͰ৳ͼΔ͕, Կ%ϓϥεʹิਖ਼͞ΕΔʁ Time ˞͜ΕҎ߱ͷσʔλ8BOUFEMZͷ͋ΔࢦඪΛͱʹՃͨ͠ͷ
©2019 Wantedly, Inc. ProphetʹΑΔ༧ଌ 'BDFCPPLͷ044ඇઢܗͳ࣌ܥྻσʔλΛقઅੑٳޮՌΛऔΓೖΕͯ༧ଌ͢Δ ɾτϨϯυ H U ɾقઅੑ T
U ɾٳɾॕޮՌ I U ɾΠϕϯτΩϟϯϖʔϯ Forecasting at Scale Sean J. Taylor∗† Facebook, Menlo Park, California, United States ࣌ܥྻσʔλ͜ΕΒͷཁૉͷͱͳ͍ͬͯΔ
©2019 Wantedly, Inc. Prophet: model࡞ Time 2016-01-01 ~ 2018-12-31·ͰͷσʔλΛֶͬͯश͠, modelΛ࡞
2019-01-01 ~ 2019-07-31ͷ༧ଌͱ࣮ࡍͷΛൺֱ͠, modelͷੑೳΛ֬ೝ͢Δ
©2019 Wantedly, Inc. Prophet: modelධՁ 2016-01-01 ~ 2018-12-31·ͰͷσʔλΛֶͬͯश͠, modelΛ࡞ 2019-01-01
~ 2019-07-31ͷ༧ଌͱ࣮ࡍͷΛൺֱ͠, modelͷੑೳΛ֬ೝ͢Δ ฏۉ4.5%ͷޡࠩ(MAPE)
©2019 Wantedly, Inc. Prophet: ֤ཁૉͷ τϨϯυ, ༵ɾ݄ͷӨڹ, ॕͷӨڹͷՄࢹԽ ॱௐʹ Լ
3্݄ঢ trend holiday weekly yearly
©2019 Wantedly, Inc. Prophet: ΠϕϯτޮՌ ɾτϨϯυ H U ɾقઅੑ T
U ɾٳɾॕޮՌ I U ɾΠϕϯτΩϟϯϖʔϯ ɾٳॕ͚ͩͰͳ͘, ҙͷΠϕϯτΩϟϯϖʔϯͷޮՌऔΓೖΕΔ͜ͱ͕Ͱ͖Δ ɾॕͷ߹ͱಉ༷ʹ, Πϕϯτ͕͋ͬͨΛࢦఆ͢Δ͚ͩͰΑ͍ ɾΠϕϯτͷʹӨڹ͕େ͖͘มΘΔͷʹద༻͠ʹ͍͔͘ɾɾɾʁ ɾ͞Βʹ, ͦΕΒͷӨڹఆྔతʹࢉग़Ͱ͖Δ
©2019 Wantedly, Inc. Prophet: ֤ཁૉͷ τϨϯυ, ༵ɾ݄ͷӨڹ, ॕͷӨڹͷՄࢹԽ ॕԼ trend
holiday weekly yearly
©2019 Wantedly, Inc. Prophet: ΠϕϯτޮՌ ɾޡࠩͷେ͖͍݄ԿΒ͔ͷΠϕϯτޮՌ͕ߟྀͰ͖͍ͯͳ͍Մೳੑ ex) ѲͰ͖͍ͯͳ͍Ωϟϯϖʔϯ͕͋ͬͨ ɾυϝΠϯࣝΛۦͯ͠ΠϕϯτޮՌΛऔΓೖΕΔͷ͕ॏཁ ࠓճհͨ͠ػೳͷ߹,
ॕ͚ͩͰతΛຬͨ͢ਫ਼͕ಘΒΕͨ
©2019 Wantedly, Inc. Prophet: ׆༻ྫ ɾࠜڌ͕໌֬ͳඪઃఆ͕Ͱ͖Δ 9݄10%৳ͼΔ͕͋ΔͷͰ, վળʹΑͬͯ5%ϓϥε͠, ߹ܭ15%৳͢ ɾ࣮ࡍͷ݁Ռ͕Prophetͷ༧ଌΛԼճͬͨ
ػೳͷ͕ಷ͖͍ͬͯͯΔ͔͠Εͳ͍ͷͰରࡦΛଧͭ ɾ9݄͕࠷͕৳ͼͦ͏ ͦͷλΠϛϯάʹ߹ΘͤͯվળࢪࡦΩϟϯϖʔϯΛߦ͏
©2019 Wantedly, Inc. ɾ࣌ܥྻσʔλͷੑ࣭ ɾɾقઅੑ(पظੑ)ɾΠϕϯτޮՌ͕͍ࠞͬͯ͟Δ ɾProphetʹΑΔ༧ଌ ɾखܰʹ࣌ܥྻ༧ଌ͕Մೳ ɾυϝΠϯ͕ࣝ͋ΕΠϕϯτޮՌͷઃఆʹΑΓਫ਼্͕͕Δ ɾσʔλΛߏ͢ΔཁૉͷӨڹ͕ఆྔతʹΘ͔Δ Summary