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
データ不足に数理モデルで立ち向かう / Japan.R 2023
Search
森下光之助
December 02, 2023
Marketing & SEO
10
5.4k
データ不足に数理モデルで立ち向かう / Japan.R 2023
2023年12月2日に行われたJapan.R 2023での発表資料です
https://japanr.connpass.com/event/302622/
森下光之助
December 02, 2023
Tweet
Share
More Decks by 森下光之助
See All by 森下光之助
baseballrによるMLBデータの抽出と階層ベイズモデルによる打率の推定 / TokyoR118
dropout009
1
530
tidymodelsによるtidyな生存時間解析 / Japan.R2024
dropout009
2
1k
回帰分析ではlm()ではなくestimatr::lm_robust()を使おう / TokyoR100
dropout009
51
10k
Counterfactual Explanationsで機械学習モデルを解釈する / TokyoR99
dropout009
3
3k
『機械学習を解釈する技術』の紹介 / Devsumi2022
dropout009
2
3.7k
シンプルな数理モデルでビジネス課題を解決する / Japan.R 2021
dropout009
2
6.3k
テレビCMのユニークリーチを最適化する / PyData.Tokyo24
dropout009
0
1.7k
Accumulated Local Effects(ALE)で機械学習モデルを解釈する / TokyoR95
dropout009
3
8.9k
データ分析手法をシミュレーションを通して理解する / stapy74
dropout009
3
1.8k
Other Decks in Marketing & SEO
See All in Marketing & SEO
The untapped power of vector embeddings
frankvandijk
0
1.1k
Perfekte Prompts für ChatGPT & Co.: Wie du KI-Tools optimal nutzt
kaispriestersbach
0
160
How to build an LLM SEO readiness audit: a practical framework
nmsamuel
1
360
What the f**k is alt-text, and how do you write it?
liamcumber
0
380
Efficient Content Optimization with Google Search Console & Apps Script
katarinadahlin
PRO
0
100
SEO Localization: From basics to best practices
raquelgonzalezseo
1
230
BrightonSEO San Diego: How to Update Content to Increase Traffic and Revenue
alyssacorso12
0
240
BrightonSEO April 2025: XML Sitemaps + Hreflang For E-Commerce
nmsamuel
1
170
A Friggin' Sweet Insider Guide to Google Sheets for SEO
greggifford
PRO
2
590
JAWS-UGとは -JAWS-UG茨城支部#0勉強会-
awsjcpm
0
130
(再)ひとり技術広報からの脱却 / Re:Breaking away from one-man technical public relations
seike460
PRO
1
230
Crafting winning internal linking strategies
smp_seo
1
460
Featured
See All Featured
Designing for humans not robots
tammielis
253
25k
Speed Design
sergeychernyshev
32
1k
We Have a Design System, Now What?
morganepeng
53
7.7k
A Modern Web Designer's Workflow
chriscoyier
695
190k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
48
2.9k
VelocityConf: Rendering Performance Case Studies
addyosmani
332
24k
Build your cross-platform service in a week with App Engine
jlugia
231
18k
Done Done
chrislema
185
16k
The World Runs on Bad Software
bkeepers
PRO
70
11k
The Myth of the Modular Monolith - Day 2 Keynote - Rails World 2024
eileencodes
26
2.9k
Music & Morning Musume
bryan
46
6.7k
Documentation Writing (for coders)
carmenintech
73
4.9k
Transcript
2023/12/02 Japan.R 2023 #JapanR @dropout009
REVISIO CDO X: @dropout009 Speaker Deck: dropout009 Blog: https://dropout009.hatenablog.com/
None
None
CM • • CM ⾒ • CM
• GRP TRP • CM ⾒ • CM ⾒ •
• CM 1 ⾒ • CM 2 1 2 3 4 5 A 1 0 1 0 1 B 0 1 0 1 0 C 1 1 1 0 1 D 0 0 1 0 0 E 0 0 0 0 0 2 (40%) 4 (80%) 7 (140%) 8 (160%) 10 (200%) 2 (40%) 3 (60%) 4 (80%) 4 (80%) 4 (80%)
• • CM × 1% 1 1
• 206% 10 2,060 69.7%
• • 0 0 頻 ⾒ 100% lm(y ~ 0
+ x) lm(y ~ 0 + log1p(x))
• •
None
l 𝑔 l CM 𝐹 Pr 𝐹 = 𝑓 ∣
𝑔 l CM 1 ⾒ 𝑟 𝑔 = Pr 𝐹 ≥ 1 ∣ 𝑔 = 1 − Pr 𝐹 = 0 ∣ 𝑔 Pr 𝐹 = 𝑓 ∣ 𝑔 𝑟 𝑔 CM
l Poisson 𝑓 𝜆 = 1 Γ 𝑓 + 1
𝜆!𝑒"# l 𝜆 𝑔 𝜆 = 𝑔 𝑟 𝑔 = 1 − Pr 𝐹 = 0 ∣ 𝑔 = 1 − 1 Γ 0 + 1 𝑔$𝑒"% = 1 − 𝑒"% dpois(f, lambda) Poisson(𝑓 ∣ 𝜆 = 5) Poisson(𝑓 ∣ 𝜆 = 3) 1 - dpois(0, g) Poisson(𝑓 ∣ 𝜆 = 2)
l 𝑟 𝑔 = 1 − 𝑒"%
l CM ⾒ CM CM CM CM Poisson(𝑓 ∣ 𝜆
= 2.06) CM
None
l CM CM CM l CM 𝜆 CM 𝜆 CM
⾒ 𝜆 Poisson(𝑓 ∣ 𝜆 = 2) Poisson(𝑓 ∣ 𝜆 = 3) Poisson(𝑓 ∣ 𝜆 = 5)
l ⾒ ⾒ 𝜆 l 𝜆 頻 𝜆 l 𝜆
Gamma 𝜆 ∣ 𝜈, 𝜈 𝜇 = 𝜈 𝜇 & Γ 𝜈 𝜆&"'𝑒" & (# E 𝜆 = 𝜇 𝜆 dgamma(nu, nu / mu) Gamma 𝜆 ∣ 1, 1 2 Gamma 𝜆 ∣ 4, 4 2 Gamma 𝜆 ∣ 16, 16 2 𝜆 𝜆
l 𝜆 ⾒ 𝜆 Pr 𝐹 = 𝑓 ∣ 𝜇,
𝜈 = ; $ ) Pr 𝐹 = 𝑓 ∣ 𝜆 𝑝 𝜆 𝜇, 𝜈 𝑑𝜆 = ; $ ) Poisson 𝑓 ∣ 𝜆 Gamma 𝜆 𝜈, 𝜈 𝜇 𝑑𝜆 = ; $ ) 1 Γ 𝑓 + 1 𝜆!𝑒"# 𝜈 𝜇 & Γ 𝜈 𝜆&"'𝑒" & (# 𝑑𝜆 = 𝜈 𝜇 & Γ 𝑓 + 1 Γ 𝜈 ; $ ) 𝜆&*!"'𝑒" &"( ( # 𝑑𝜆 = 𝜈 𝜇 & Γ 𝑓 + 1 Γ 𝜈 Γ 𝜈 + 𝑓 𝜈 + 𝜇 𝜇 &*! ; $ ) 𝜈 + 𝜇 𝜇 &*! Γ 𝜈 + 𝑓 𝜆&*!"'𝑒" &*( ( # 𝑑𝜆 = Γ 𝜈 + 𝑓 Γ 𝑓 + 1 Γ 𝜈 𝜈 𝜈 + 𝜇 & 𝜇 𝜈 + 𝜇 ! = , ! " Gamma 𝜆 𝜈 + 𝑓, 𝜈 + 𝜇 𝜇 𝑑𝜆 = 1
l ⾒ Negative Binomial Distribution; NBD NB 𝑓 𝜇, 𝜈
= Γ 𝜈 + 𝑓 Γ 𝑓 + 1 Γ 𝜈 𝜈 𝜈 + 𝜇 & 𝜇 𝜈 + 𝜇 ! NB 𝑓 2.06,1 NB 𝑓 2.06,3 NB 𝑓 2.06,10 dnbinom(f, mu = mu, size = nu)
l ⾒ 𝑟 𝑔, 𝜈 = 1 − Pr 𝐹
= 0 ∣ 𝑔, 𝜈 = 1 − Γ 𝜈 + 0 Γ 0 + 1 Γ 𝜈 𝜈 𝜈 + 𝑔 & 𝑔 𝜈 + 𝑔 $ = 1 − 𝜈 𝜈 + 𝑔 & l 𝜈 1 - dnbinom(0, mu = g, size = nu) 𝑟 𝑔, 1 𝑟 𝑔, 3 𝑟 𝑔, 10
l 𝑟 𝑔, 𝜈 𝜈 𝜈 l 𝑟+ 𝑔+ ̂
𝜈 ̂ 𝜈 = argmin & 1 − 𝜈 𝜈 + 𝑔+ & − 𝑟′ l ̂ 𝜈 𝑟 𝑔, ̂ 𝜈 = 1 − ̂ 𝜈 ̂ 𝜈 + 𝑔 , & CM CM
l 1 ⾒ CM 3 CM ⾒ l CM 𝑓
⾒ 𝑓 + l 𝑓 + 𝑟!* 𝑔, 𝜈 = Pr 𝐹 ≥ 𝑓 ∣ 𝑔, 𝑣 = 1 − Pr 𝐹 ≤ 𝑓 − 1 ∣ 𝑔, 𝜈 = 1 − E !!-$ !"' Γ 𝜈 + 𝑓+ Γ 𝑓+ + 1 Γ 𝜈 𝜈 𝜈 + 𝑔 & 𝑔 𝜈 + 𝑔 !! 𝑓 𝑓 + 𝑟!" 𝑟#" 𝑟$" 1 - pnbinom(f - 1, mu = g, size = nu)
None
l l l l ⾒
• Goerg, Georg M. "Estimating reach curves from one data
point." (2014).