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
CTRオンライン予測システムのアーキテクチャ
Search
Yoshitomo Hayashi
September 27, 2017
Technology
4.5k
2
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
CTRオンライン予測システムのアーキテクチャ
「Ameba広告システムの裏側見せます - オレシカナイトvol3」発表資料です。
https://cyberagent.connpass.com/event/64176/
Yoshitomo Hayashi
September 27, 2017
More Decks by Yoshitomo Hayashi
See All by Yoshitomo Hayashi
Ameba DSPのOpen-Auctionにおける入札戦略
yyhayashi303
2
3k
進化する配信ロジックとDSP戦略
yyhayashi303
1
170
モブプロ導入で見えてきた効果@オレシカナイト
yyhayashi303
1
1.2k
CircuitBreakerの適用
yyhayashi303
0
1.8k
Other Decks in Technology
See All in Technology
会社紹介資料 / Sansan Company Profile
sansan33
PRO
24
430k
AIにブラウザを触らせて、E2EテストをPlaywrightで書く
koji_kawamura
0
250
LLM Internals: 언어 모델의 계보와 알고리즘 진화 (2023~2026)
inureyes
PRO
1
830
関東Kaggler会発表資料
takoi
1
270
ガバメント AI 源内を地方自治体は活用できるのか可能性と課題、期待について
takeda_h
2
460
Go 1.27 の標準パッケージに uuid が入った!のでいろいろ喋る / go_127_std_go_uuid
convto
3
560
Rust×eBPFでEDRっぽいものをつくる
sunlife3
2
670
DDDのエッセンスを取り入れたAIでの開発
ak2ie
0
190
Kiro Crew入門 - 常駐エージェントの仕組みと使いどころ / Intro to Kiro Crew
k_adachi_01
1
700
英語が話せなくてもKubeConスタッフに参加した話
yusuke427
1
1.1k
老害フォレンジッカーはAI羊の夢を見るか?
tadmaddad
0
360
【GCC2026】TrueHDRIを用いたルックデブ環境とライティングテクニック
bandainamcostudios
PRO
0
150
Featured
See All Featured
Max Prin - Stacking Signals: How International SEO Comes Together (And Falls Apart)
techseoconnect
PRO
0
420
The Art of Programming - Codeland 2020
erikaheidi
57
14k
svc-hook: hooking system calls on ARM64 by binary rewriting
retrage
2
510
What the history of the web can teach us about the future of AI
inesmontani
PRO
1
660
Google's AI Overviews - The New Search
badams
0
1.1k
Skip the Path - Find Your Career Trail
mkilby
1
190
Fireside Chat
paigeccino
42
4k
We Have a Design System, Now What?
morganepeng
55
8.3k
How Software Deployment tools have changed in the past 20 years
geshan
1
34k
How to Get Subject Matter Experts Bought In and Actively Contributing to SEO & PR Initiatives.
livdayseo
0
180
Navigating the moral maze — ethical principles for Al-driven product design
skipperchong
2
500
Stop Working from a Prison Cell
hatefulcrawdad
274
21k
Transcript
CTRΦϯϥΠϯ༧ଌ γεςϜͷΞʔΩςΫνϟ ΦϨγΧφΠτvol.3 גࣜձࣾαΠόʔΤʔδΣϯτ ྛ ۍ๎
ࣗݾհ ྛɹۍ๎ʢ͠Α͠ͱʣ αʔόʔαΠυΤϯδχΞ 2011 αΠόʔΤʔδΣϯτೖࣾ ೖࣾޙ͍͔ͭ͘ͷwebαʔϏε୲ 2015͔ΒMDHͰ৴ϩδοΫͷվળ
ࣗݾհ झຯ ίʔώʔ ՈͰυϦοϓͨ͠Γ ίʔϧυϒϦϡʔ࡞ͬͨΓ Netflix Ξχϝ ΥʔΩϯάɾσου
CTRΦϯϥΠϯ༧ଌ γεςϜͷΞʔΩςΫνϟ
CTRͱʁ
CTRͱ? CTRʢΫϦοΫʣ CTR = Click / Impression
ͳͥ༧ଌ͕ඞཁͳͷ͔ʁ
ࠂͷऩӹ ΫϦοΫ՝ۚ ࠂ1ΫϦοΫ͋ͨΓʓʓԁʢCPCʣ
ࠂͷऩӹ ΫϦοΫ՝ۚ ࠂ1ΫϦοΫ͋ͨΓʓʓԁʢCPCʣ ΑΓΫϦοΫ͞ΕΔ CPC͕ߴ͍ࠂΛ৴͍ͨ͠
ࠂΛϥϯΩϯά CTR × CPCʹΑΔϥϯΩϯά CTRόϯσΟοτΞϧΰϦζϜͰࢉग़ ࣮ͷimp, clickΛ༻ ࣮ͷूܭ [ࠂ ×
Ϣʔβʔଐੑ × ࠂ] ຖʹimp, clickΛܭଌ
͜Ε·ͰͷϥϯΩϯάͷ՝ ࣮͕গͳ͍߹ʹ͏·͘࠷దԽ͞Εͳ͍ Ϣʔβʔຖʹ࠷దͳࠂΛ৴ग़དྷ͍ͯͳ͍
ػցֶशʹΑΔCTR༧ଌ ଟ͘ͷૉੑΛՃՄೳ ࣮͕গͳͯ͋͘Δఔ༧ଌՄೳ ύʔιφϥΠζ͞ΕͨϥϯΩϯά
։ൃମ੍ ॳظ γεςϜΤϯδχΞ2ਓ σʔλαΠΤϯςΟετ1ਓʢळ༿ݪϥϘʣ ݱࡏ γεςϜΤϯδχΞ2ʙ3ਓ σʔλαΠΤϯςΟετ2ਓʢळ༿ݪϥϘ1ਓʣ
ΞʔΩςΫνϟ
ΞʔΩςΫνϟ ֶशσʔλ JoinࡁΈͷ ϩά ϩά ϩά ػցֶश ϝτϦΫε Ϟσϧ ϑΝΠϧ
Ϟσϧ ϑΝΠϧ Stream Aggregator Data Joiner Learner Predictor ModelStore
Data-Joiner
Data-Joinerͷׂ ࠂ͕click͞Ε͔ͨͲ͏͔Λఆ͢Δ imp౸ୡޙɺҰఆظؒclickͷ౸ୡΛͭ impͱclickͷσʔλͷࠩҟΛແ͘͢ refererͷʹࠩҟ͕͋ͬͨ
Apache IgniteΛ༻࣮ͨ͠ Event Notifications imp click Stream Data Joiner Aggregator
Apache IgniteΛ༻࣮ͨ͠ Event Notifications put-event impͷ߹ TTLΛ5ʹઃఆ͠อଘ imp click Stream
Data Joiner Aggregator
Apache IgniteΛ༻࣮ͨ͠ Event Notifications put-event impͷ߹ TTLΛ5ʹઃఆ͠อଘ put-event clickͷ߹ରԠ͢Δ impΛݕࡧ͠click͞Εͨ
impͱͯ͠Aggregatorʹૹ৴ imp click Stream Data Joiner Aggregator
Apache IgniteΛ༻࣮ͨ͠ Event Notifications put-event impͷ߹ TTLΛ5ʹઃఆ͠อଘ put-event clickͷ߹ରԠ͢Δ impΛݕࡧ͠click͞Εͨ
impͱͯ͠Aggregatorʹૹ৴ expired-event 5ܦͬͯclick͕དྷͳ͍ ߹click͞Εͳ͔ͬͨimpͱ ͯ͠Aggregatorʹૹ৴ imp click Stream Data Joiner Aggregator
Aggregator
Aggregatorͷׂ ֶशʹඞཁͳσʔλΛऩू ϩά͔Βऔಘग़དྷΔͷ ϩά͔ΒऔಘͰ͖ͳ͍ͷ ऩूͨ͠σʔλΛLearnerʹૹΔ σʔλαϯϓϦϯά͠100݅ͣͭ
ϩά͔ΒऔಘՄೳͳσʔλ Ϣʔβʔͷใ UA, IPΞυϨε, referer, ϦλήϢʔβʔ͔ etc. ࠂͷใ ۀछID, ࠂओID,
ΫϦΤΠςΟϒID etc. ໘ͷใ ࠂID, ϝσΟΞID etc.
ϩά͔ΒऔಘෆՄೳͳσʔλ Ϣʔβʔͷใ σϞάϥʢੑผɾྸʣ ಠࣗʹ࡞ͨ͠ૉੑ ߦಈཤྺΛϕʔεʹͨ͠ϢʔβʔΫϥελ ࠂςΩετΫϥελ
Learner
Learnerͷׂ Aggregator͔Βड͚औͬͨσʔλΛ༻͠ϞσϧΛߏங͢Δ ϞσϧΛߏங͢ΔίΞ෦C++ʢσʔλαΠΤϯςΟ ετ୲ʣ 15ִؒͰϞσϧΛߋ৽ ϞσϧϑΝΠϧͰϞσϧετΞʢS3ʣʹอଘ ػցֶशϝτϦΫεʢauroc, logloss etc.ʣΛDBʹอଘ
Learner͕ੜ͢Δ ϞσϧϑΝΠϧ ֶश༻ϚελϑΝΠϧʢ12ִ࣌ؒؒͰߋ৽ʣ Learner͜ͷϑΝΠϧΛ༻ֶͯ͠श͢Δ ༧ଌ༻ϚελϑΝΠϧʢ12ִ࣌ؒؒͰߋ৽ʣ Predictor͕CTRΛ༧ଌ͢Δࡍʹ༻͢Δ ༧ଌ༻ύονϑΝΠϧ ༧ଌ༻ϚελϑΝΠϧ͔ΒͷࠩϑΝΠϧ ͜ͷϑΝΠϧ͕LearnerʹΑͬͯ15ִؒͰߋ৽͞ΕΔ
Predictor
Predictor ༧ଌʹඞཁͳσʔλΛड͚औΓ༧ଌCTRΛฦ͢API 15ִؒͰϞσϧετΞ͔Β࠷৽ͷϞσϧϑΝΠ ϧΛऔಘ͠ߋ৽͢Δ
༧ଌ͕։࢝ग़དྷΔ·Ͱ ىಈ࣌ʹϚελϑΝΠϧΛಡΈࠐΉ ϚελϑΝΠϧಡΈࠐΈޙ͔Β༧ଌՄೳ ϚελϑΝΠϧͷ࠷ऴߋ৽͔࣌Βͷࠩύον ϑΝΠϧΛద༻͍ͯ͘͠ʢ࠷େͰ47ϑΝΠϧʣ 1ύονϑΝΠϧ5ʙ6ඵͰద༻ՄೳͳͷͰɺ5 ҎͰྃ
CTR༧ଌγεςϜಋೖޙ
CPMൺֱ ϥϯμϜ ࣮CTR ༧ଌCTR ࣮CTRରൺ7ˋվળ ϥϯμϜରൺ50ˋվળ
CTRൺֱ ࣮CTRରൺ6ˋվળ ϥϯμϜରൺ8ˋվળ ϥϯμϜ ࣮CTR ༧ଌCTR
͓·͚
Real-Time Optimizer ʢROʣ
Real- Time OptimizerʢROʣ 20176݄ϦϦʔε ඪCPAʹ߹ΘͤͯCPCΛࣗಈௐ͢Δ
CPCͷࢉग़ํ๏ CPC = ඪCPA × CVR ඪCPA 1CVಘΔͷʹࢧͬͯྑ͍අ༻ ࠂओ͕ܾఆ CVRʢίϯόʔδϣϯʣ
CV / Click
͜Ε·ͰͷCVRࢉग़ํ๏ աڈ࣮ΛݩʹώϡʔϦεςΟοΫͳCVRࢉग़ ࠂ × ΞΧϯτຖʹCVRΛࢉग़ ϦλήϢʔβʔͷ߹ิਖ਼Ϩʔτ
ݱঢ়ͷ՝ CV͕ग़ʹ͍͘Ҋ݅ͷ߹ʹCVRͷਫ਼͕ѱ͍ ৽نΞΧϯτͷ߹࣮͕ແ͍ ۀछશମͷ࣮Λࢀর ৽نࠂͷ߹ʹ࣮͕ͳ͍ ࠂάϧʔϓͷ࣮Λࢀর
CTR༧ଌγεςϜͷࢿ࢈Λ ׆༻ͨ͠CVR༧ଌγεςϜ
CVR༧ଌγεςϜ Ϟσϧͷߏங࣍όονʢBatch-Learnerʣ Clickൃੜ͔ΒCVൃੜ·Ͱͷ͕͍࣌ؒͷͰϦΞ ϧλΠϜͰֶश͠ͳ͍ LearnerͷίΞ෦ɺੜ͞ΕΔϞσϧϑΝΠ ϧɺPredictorͦͷ··༻
ΞʔΩςΫνϟ Predictor ModelStore Batch-Learner ϞσϧϑΝΠϧ ϞσϧϑΝΠϧ ػցֶश ϝτϦΫε ֶशσʔλ
CVR༧ଌͷঢ়گ ༧ଌCVR ࣮CVR ࣮CVRରൺ3ˋվળ
͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠