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
Real-Time_Bidding_Algorithms_for_performance-Ba...
Search
jujudubai
August 17, 2014
Research
0
730
Real-Time_Bidding_Algorithms_for_performance-Based_Display_Ad_Allocation.pdf
いろいろと参考にしながら、要約を。
この論文はとても参考になります。
jujudubai
August 17, 2014
Tweet
Share
More Decks by jujudubai
See All by jujudubai
juju1008
juju1008
1
4.2k
Realtime Bid Optimization with Smooth Budget Delivery in Online Advertising
juju1008
2
900
Estimating Conversion Rate in Display Advertising from Past Performance Data
juju1008
1
890
Other Decks in Research
See All in Research
SpectralMamba: Efficient Mamba for Hyperspectral Image Classification
satai
3
200
Sosiaalisen median katsaus 03/2025 + tekoäly
hponka
0
490
SI-D案内資料_京都文教大学
ryojitakeuchi1116
0
230
Weekly AI Agents News! 12月号 プロダクト/ニュースのアーカイブ
masatoto
0
360
JSAI NeurIPS 2024 参加報告会(AI アライメント)
akifumi_wachi
5
930
知識強化言語モデルLUKE @ LUKEミートアップ
ikuyamada
0
340
Introduction of NII S. Koyama's Lab (AY2025)
skoyamalab
0
160
さくらインターネット研究所 アップデート2025年
matsumoto_r
PRO
0
430
LLM 시대의 Compliance: Safety & Security
huffon
0
630
NLP2025SharedTask翻訳部門
moriokataku
0
260
PostgreSQLにおける分散トレーシングの現在 - 第50回PostgreSQLアンカンファレンス
seinoyu
0
290
言語モデルの内部機序:解析と解釈
eumesy
PRO
32
13k
Featured
See All Featured
How STYLIGHT went responsive
nonsquared
99
5.4k
A Philosophy of Restraint
colly
203
16k
Building Adaptive Systems
keathley
41
2.5k
A Modern Web Designer's Workflow
chriscoyier
693
190k
GraphQLの誤解/rethinking-graphql
sonatard
70
10k
Building an army of robots
kneath
304
45k
Thoughts on Productivity
jonyablonski
69
4.5k
The Invisible Side of Design
smashingmag
299
50k
Docker and Python
trallard
44
3.3k
The Cult of Friendly URLs
andyhume
78
6.3k
Intergalactic Javascript Robots from Outer Space
tanoku
270
27k
Building a Modern Day E-commerce SEO Strategy
aleyda
39
7.2k
Transcript
Review: “Real-Time Bidding Algorithms for performance-Based Display Ad Allocation” Tatsuki
Sugio
ຊจͷ֓ཁ A. demand-side, supply-side • ༧ࢉࢿͷ࠷దԽɺऩӹʢrevenueʣͷ࠷େԽ • RTB Exchangeʹ͓͍ͯɺimpຖʹΩϟϯϖʔϯΛׂΓͯΔ ➡
ϦΞϧλΠϜͰͷ࠷దԽʹΑΓ࣮ݱ ➡ errorͷେ͖͞ʹԠͯ͡ύϥϝʔλΛௐ B. ՝ • มɺ੍͕ଟ͍ ➡ ઢܗܭըͷରͷղʹΑΓ࣮ݱ • ΦϑϥΠϯ࠷దԽͰཻ͕ૈ͍ ࢢͷมԽʹରͯ͠దԠతͳbid͕Ͱ͖ͳ͍ ➡ ϦΞϧλΠϜͰͷ࠷దԽʹΑΔࡉཻ͔͍Ͱͷ࠷దԽΛ࣮ݱ C. ํ๏ • online bidding algorithm frameworkΛఏҊ • Ωϟϯϖʔϯຖͷbidػೳύϥϝʔλͷߋ৽ํ๏ʢWaterlevel or Model-based ʣͱͯ͠ɺطଘͷϦιʔε ͷۙࣅΞϧΰϦζϜʹinspire͞Εͨํ๏ͱɺbidͷউͷΛϞσϧԽͯࣜ͠ʹΈࠐΜͩͷΛఏҊɻ
Formulation A. ऩӹͷఆٛ B. ೖࡳֹͷܾఆɺௐ ࠂओผ
ೖࡳֹௐͷ߲ ͜Ε͔Β͜ͷzЋzΛٻΊͯɺ࠷దͳzCJEQSJDFzΛਪఆ͠·͢
LR Formulation • ࠷దԽ ΩϟϯϖʔϯKͷJ൪ͷJNQνϟϯεʹJNQͰ͖͔ͨ൱͔ʢೋʣ WJKQJK RJKˡ $53 $1$ ΩϟϯϖʔϯKͷඪJNQʢ༧ࢉ੍Λ݉ͶΔʣ
εϥοΫ݅
• ࠷దԽͷର
➡ α,βΛٻΊΔ͜ͱ͕త ܭࢉճɺO(mn)Ͱͳ͘ɺO(m+n) ➡ શϢχϞδϡϥߦྻʢtotally unimodular matrix, TU ߦྻʣʹجͮ͘ ࢀߟʣhttp://ja.wikipedia.org/ ๚ऀͷ૿ՃͷܦࡁతʢJNQͷ࠷খՁ֨ͱʣ ༧ࢉͷ૿Ճͷܦࡁతʢ࠷খརӹͱʣ
Real-Time Bidding Algorithm • ٙࣅίʔυ HPBMBDIJFWFE Ќͷܭࢉ POMJOF"MHPSJUINͷద༻
Control-theoretic Bid Adjustment • waterlevel-base update (online algorithm) - ίετߟྀ͠ͳ͍
- PIɺPIDཧ JNQ FSSPS FSSPSʹͲΕ͚ͩૣ͘Ԡ͢Δ͔ͷ
1*%੍ޚཧ 1*%੍ޚͷجຊࣜɺภࠩFʹൺྫ͢Δग़ྗΛग़͢ൺྫಈ࡞ʢ1PQPSUJOBMBDUJPO1ಈ࡞ʣͱɺ ภࠩFͷੵʹൺྫ͢Δग़ྗΛग़͢ੵಈ࡞ʢ*OUFHSBMBDUJPO*ಈ࡞ʣͱɺ ภࠩFͷඍʹൺྫ͢Δग़ྗΛग़͢ඍಈ࡞ʢ%FSJWBUJWFBDUJPO%ಈ࡞ʣ͔ΒͳΔɻ ௨ৗɺ1ಈ࡞Λओମʹͯ͠ɺิॿతʹ*ಈ࡞ͱ%ಈ࡞Λ੍ޚରʹԠͯ͡దʹΈ߹ΘͤΔɻ ૢ࡞ྔ.7ɺͦΕͧΕͷͱͯ͠ɺ࣍ࣜͷ༷ʹද͞ΕΔɻ IUUQXXXOJDPNXIJUFQBQFSKB
Model-based Bid Adjustment • γεςϜ੍ޚཧʹجͮ͘Ξϓϩʔν(PI:online algorithm) - ίετɺೖࡳֹߟྀ FSSPSʹૣ͘ͲΕ͚ͩૣ͘Ԡ͢Δ͔ͷ ཧతͳೖࡳՁ֨
ཧతͳউʢHJʹ߹ΘͤΔͨΊʹඞཁͳউʣ ؍ଌ͞Εͨউ ೖࡳίετ .-&ͷύϥϝʔλɻ XJOͨ͠ೖࡳ X ͷ౷ܭྔ͔Βಋ͔ΕΔɻ
a Practical formulation • ίετ߲ͷಋೖʹΑΓߋʹҰൠԽͨ͠ओ
• ίετ߲ͷಋೖʹΑΓߋʹҰൠԽͨ͠ର JNQ(SPVQ QMBDFNFOU Jͷ֫ಘͰ͖ͦ͏ͳJNQ
Experiments • ࣮ݧ݁Ռͷ֓ཁ - αͷௐʹΑͬͯೖࡳͷ࠷దԽ͕ߦ͑Δ͔Ͳ͏͔ - ҟͳΔ࠷దԽख๏ͷಋೖʹΑΓͲͷఔύϑΥʔϚϯε͕ҟͳΔͷ͔ - αͷॳظ͕ͲͷఔӨڹ͢Δͷ͔ •
࣮ݧ݅ - ༻σʔλσΟεϓϨΠωοτϫʔΫͷσʔλ - ฏۉ120Mͷimp͕͋ΔαΠτͰ࣮ݧ - 4ͭͷCPCΩϟϯϖʔϯ͕ର • σʔλ • timestamp,placement,user,campaign,clicks,impressions • ॱʹt,i=(placement:user),j,cij(t),xij(t)
MJGU ʹ ࢪࡦΛ࣮ࢪ͠ͳ͍࣌ͷ݁Ռ ࢪࡦΛ࣮ࢪͨ࣌͠ͷ݁Ռ IUUQXXXBMCFSUDPKQUFDIOPMPHZDSNMJGUIUNM
- Experiments 1 • ؍ଌͱγϡϛϨʔγϣϯʹΑΔͷlift ➡ offlineͷΈΑΓonlineͰαΛௐͨ͠ํ͕͕ྑ͍
➡ model-based bid ͱ Waterlevel bidͷൺֱ - offlineͰͷαͷࢉग़1ͷσʔλ - αࢉग़ޙͷ4ؒͷσʔλΛൺֱ ➡ online algorithmoffline algorithmʹରͯ͠90ˋҎ্ͷ ➡ ҆ఆੑModel Bidder͕ྑ͍
- Experiments 2 • hourlyͷมಈʢ࣌ؒͷ҆ఆੑ֬ೝʣ ➡ Waterlevel Bidder࣌ؒతͳ҆ఆੑ͕ߴ͍ ➡ Model
Bidderෆ҆ఆ
- Experiment 3 • online algorithm(Waterlevel Bidder)ʹ͓͚ΔαͷॳظͷӨڹ ➡ ॳظͷมಈ΄ͱΜͲͳ͍ ͔͠͠ɺΩϟϯϖʔϯ༧ࢉͷ੍͕ݫ͚͠ΕӨڹ͕͋Δ͔…
• ༧ࢉ੍ʢݫʣ ➡ ༧ࢉ੍͕ݫ͚͠Εɺ ॳظͷมಈ͋Δɻ offline࠷దԽͨ͠αͷ͕ྑ͍ɻ - Experiments 4
Conclusion • ݁ - γϯϓϧ͕ͩཧతഎܠͷ͋Δonline algorithmΛఏҊ - PIDཧͷԠ༻Մೳੑ - ଞͷछྨͷϞσϧߟྀ͢Εɺߋʹվྑ͕ग़དྷΔͷͰͳ͍͔