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を使った時系列データ予測と機械学習モデルとの比較 / prophet-vs-ml
Search
Haruki Okuyama
October 27, 2019
Business
0
1.3k
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
時系列データにおいて, Prophetと3時間で作成した機械学習モデルとの精度比較
Haruki Okuyama
October 27, 2019
Tweet
Share
More Decks by Haruki Okuyama
See All by Haruki Okuyama
Prophetを使ったコスパの良い時系列データ予測 / prophet-use-cases
spring1018
0
110
Other Decks in Business
See All in Business
Kyash TechTalk #8 Kyashにおけるクレジット事業部とは
sayueda
0
130
2025.10_中途採用資料.pdf
superstudio
PRO
2
84k
AIプロダクト時代のPdMに必要なコンテキストエンジニアリング
kayato
0
110
テオリア・テクノロジーズ:About Us
theoriatec2024
2
39k
ソニックガーデン経営組織論(2025/10版)
kuranuki
1
2.4k
Recept_Culture deck
shokennakase
0
180
株式会社STYZ会社概要資料
culumu
0
370
採用ピッチ資料
awesome22
0
310
Algomatic Works Company Deck
algomatic
PRO
1
760
株式会社SAFELY 会社紹介 / Company
safely_pr
1
4.2k
株式会社Nagisa_カンパニーデック
nagisa_recruit
0
360
THxApp project
shotaro0913
0
450
Featured
See All Featured
Imperfection Machines: The Place of Print at Facebook
scottboms
269
13k
Docker and Python
trallard
46
3.6k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
140
34k
Typedesign – Prime Four
hannesfritz
42
2.8k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
31
2.7k
Bootstrapping a Software Product
garrettdimon
PRO
307
110k
The Myth of the Modular Monolith - Day 2 Keynote - Rails World 2024
eileencodes
26
3.1k
Done Done
chrislema
185
16k
Java REST API Framework Comparison - PWX 2021
mraible
34
8.9k
Context Engineering - Making Every Token Count
addyosmani
7
280
Into the Great Unknown - MozCon
thekraken
40
2.1k
Rebuilding a faster, lazier Slack
samanthasiow
84
9.2k
Transcript
©2019 Wantedly, Inc. ProphetΛͬͨ࣌ܥྻσʔλ༧ଌͱ ػցֶशϞσϧͱͷൺֱ ML for Beginners! MeetUp #1
LTձ Oct 27, 2019 - Haruki Okuyama - @spring1018x
©2019 Wantedly, Inc. Self-Introduction •Haruki OkuyamaʢԞࢁ ݰكʣ •Chemistry Research (until
March 2019) •Wantedly, Inc. (since April 2019) •Recommendation Team • Mainly, Data Analysis
©2019 Wantedly, Inc. ͢͜ͱ ɾ౷ܭϞσϧͱػցֶशϞσϧͷ࣌ܥྻσʔλ༧ଌ ɾ࣌ܥྻ༧ଌϥΠϒϥϦProphetͷհ ͞ͳ͍͜ͱ ɾProphetͷৄࡉͳΞϧΰϦζϜ About this
talk
©2019 Wantedly, Inc. ɾֶशσʔλ(աڈ)ͱςετσʔλ(ະདྷ)ͷ͕ҟͳΔ ɾτϨϯυ͕͋Δͱ2ͭͷ͕ҟͳΔͷવ ɾػցֶशϞσϧΑΓ౷ܭϞσϧͷํ͕༧ଌਫ਼͕͍͍߹͋ͬͨΓ ͢ΔΒ͍͠* => ࣌ܥྻσʔλʹରͯ͠ػցֶशϞσϧͱ౷ܭϞσϧͷ༧ଌ݁ՌΛ ൺֱ͠,
ײ৮Λ͔֬Ί͍ͨʂ ࣌ܥྻσʔλͷ༧ଌͷ͠͞ *https://tjo.hatenablog.com/entry/2019/09/18/190000 https://t.co/S3BpRgtxUW?amp=1
©2019 Wantedly, Inc. ɾػցֶशϞσϧ ϥάಛྔΛத৺ʹ15ݸͷಛྔΛ࡞͠, LightGBMͰֶश ɾ౷ܭϞσϧ FacebookOSSͷProphetΛ༻ : ͋Δࢦඪ͕དྷ݄͍ͭ͘ʹͳΔ͔Λ༧ଌ͢Δ
* ͋Δͷ࣌Ͱ࣍ͷ݄ͷࢦඪΛ༧ଌ͍ͨ͠ ** 3࣌ؒ͘Β͍Ͱ༧ଌϞσϧΛ࡞͍ͨ͠
©2019 Wantedly, Inc. Prophetͬͯʁ 'BDFCPPLͷ044ඇઢܗͳ࣌ܥྻσʔλΛقઅੑٳޮՌΛऔΓೖΕͯ༧ଌ͢Δ ɾτϨϯυ g(t) ɾقઅੑ(पظੑ) s(t) ɾٳɾॕޮՌ
h(t) ɾΠϕϯτΩϟϯϖʔϯ Forecasting at Scale Sean J. Taylor∗† Facebook, Menlo Park, California, United States ࣌ܥྻσʔλΛ͜ΕΒͷཁૉͷͱߟ͑Δ (*ࣗݾ૬ؔߟྀ͍ͯ͠ͳ͍)
©2019 Wantedly, Inc. model࡞ Time 2016-01-01 ~ 2019-06-30·ͰͷσʔλΛֶͬͯश͠, modelΛ࡞ 2019-07-01
~ 2019-09-30ͰධՁ
©2019 Wantedly, Inc. modelධՁ Prophetͷํ͕ਫ਼͕ߴ͍ 2016-01-01 ~ 2019-06-30·ͰͷσʔλΛֶͬͯश͠, modelΛ࡞ 2019-07-01
~ 2019-09-30ͰධՁ Prophet ML MAPE: 12.2% MAPE: 10.6%
©2019 Wantedly, Inc. ͜͜·Ͱͷ·ͱΊ ɾτϨϯυɾपظੑͷڧ͍࣌ܥྻσʔλʹରͯ͠, ػցֶशϞσϧͷ߹, ࣌ؒͰਫ਼Λग़͢ͷ͍͠ ɾͬͱ࣌ؒΛ͔͚ΒΕΔ߹ಛྔઃܭɾόϦσʔγϣϯͷͷ༨ ͕͋ΔͷͰ࣌ܥྻϞσϧʹউͯͦ͏
©2019 Wantedly, Inc. Prophet: ֤ཁૉͷӨڹ ɾτϨϯυ g(t) ɾقઅੑ s(t) ɾٳɾॕޮՌ
h(t) ɾΠϕϯτΩϟϯϖʔϯ ɾٳॕ͚ͩͰͳ͘, ҙͷΠϕϯτΩϟϯϖʔϯͷޮՌऔΓೖΕΔ͜ͱ͕Ͱ͖Δ ɾ͞Βʹ, ֤ཁૉͷӨڹఆྔతʹࢉग़Ͱ͖Δ
©2019 Wantedly, Inc. Prophet: ֤ཁૉͷӨڹ τϨϯυ, ༵ɾ݄ͷӨڹ, ॕͷӨڹͷՄࢹԽ ॱௐʹ Լ
3্݄ঢ trend holiday weekly yearly ॕԼ
©2019 Wantedly, Inc. Prophetͷಛ ɾτϨϯυɾपظੑͷڧ͍࣌ܥྻσʔλʹ͍͍ͯΔ ɾ֤ͷӨڹ͕ఆྔతʹѲͰ͖ΔͷͰઆ໌ੑ͕ߴ͍ ɾύϥϝʔλνϡʔχϯάͦΕ΄Ͳඞཁͳ͍ ɾతͱ͢ΔࢦඪͷυϝΠϯࣝ͑͋͞ΕΑ͍. ྫ༵͑पظ͕͋Δ, ࿈ٳԼ͢Δetc
©2019 Wantedly, Inc. ɾ࣌ܥྻσʔλ༧ଌʹ͓͚ΔػցֶशϞσϧͱProphetͷൺֱ ɾઃఆ͔͚ΒΕΔ࣌ؒΛߟྀͯ͠બ͢Δ ɾProphetʹΑΔ༧ଌ ɾखܰʹ࣌ܥྻ༧ଌ͕Մೳ ɾυϝΠϯ͕ࣝ͋ΕपظɾΠϕϯτޮՌͷઃఆʹΑΓਫ਼্͕͕Δ ɾσʔλΛߏ͢ΔཁૉͷӨڹ͕ఆྔతʹΘ͔Δ Summary