オンライン広告関連の論文を50本くらい雑に紹介する AdKDD編 / adkdd-all

ΦϯϥΠϯ޿ࠂؔ࿈ͷ࿦จΛ
50ຊ͘Β͍ࡶʹ঺հ͢Δ
AdKDDฤ
@todesking

View Slide

ࣗݾ঺հ
• @todesking

• DSP/ΞυωοτϫʔΫͷձࣾ
ͰػցֶशΛ΍͍ͬͯΔ

• ೔ࠒͷٕज़ௐ͕ࠪ଍Γͳ͍ͳ
ͱࢥͬͨͷͰAdKDDͷ࿦จΛ
ோΊͯΈ·ͨ͠

View Slide

AdKDD/TargetADͱ͸
• KDD: Knowledge Discovery and Data Mining

• σʔλϚΠχϯά΍ػցֶशʹؔ͢ΔԠ༻తͳ࿩୊Λѻ͏ࠃࡍ
ձٞ

• AdKDD: ΦϯϥΠϯ޿ࠂؔ࿈ٕज़Λѻ͏෼Պձ

• TargetAd: WSDMͱ͍͏ࠃࡍձٞʹ͓͍ͯΦϯϥΠϯ޿ࠂؔ࿈ٕज़
Λѻ͏෼Պձ

• ࠷ۙ͸AdKDDʹٵऩ͞Εͨ?

• ࠾୒࿦จ਺͸ຖ೥10ຊऑ

• KDDຊମʹ΋޿ࠂܥ࿦จ͕͋Δ(੗Έ෼͚͸Ṗ)

View Slide

஫ҙ
• ཁ໿͸΄΅AbstractோΊ͚ͨͩɺਖ਼֬͞͸อূ͠·ͤΜ

• ؒҧ͍ͷࢦఠ͸׻ܴ

• ۙ೥ͷ࿦จ͸ެࣜαΠτ͔ΒPDFͷ௚ϦϯΫ͕͋Δ

• Keynote/Invited talk͸লུ

View Slide

AdKDD 2014
https://dblp.org/db/conf/kdd/adkdd2014

View Slide

A Dynamic Pricing Model for Unifying
Programmatic Guarantee and Real-Time
Bidding in Display Advertising
• publisher͕impΛചΔʹ͸RTBͱguranteed contractͱ͍
͏ೋͭͷํ๏͕͋Δ

• publisherͷऩӹΛදݱ͢Δ਺ཧϞσϧΛ࡞Γɺೋͭͷํ
๏΁ͷ༧ࢉ഑෼Λ࠷దԽ͢Δ

View Slide

Multi-Touch Attribution Based Budget
Allocation in Online Advertising
• Ωϟϯϖʔϯશମͷ༧ࢉ͕ܾ·͓ͬͯΓɺԼҐͷΩϟϯ
ϖʔϯʹ഑෼͍ͨ͠

• ໨ඪ΁ͷߩݙʹԠͯ͡഑෼͍ͨ͠ˠlast-touch/multi-
touch attributionͷ৘ใΛݩʹׂͯ͠Γ౰ͯΔํ๏Λ࣮૷
ͨ͠

View Slide

Pleasing the advertising oracle:
Probabilistic prediction from sampled,
aggregated ground truth
• ޿ࠂΩϟϯϖʔϯͷηάϝϯτਫ਼౓Λୈࡾऀ͕ධՁ͢Δ
έʔε

• ධՁͱͯ͠ɺ޿ࠂΛݟͨϢʔβͷσϞάϥଐੑͷׂ߹͕
ϑΟʔυόοΫ͞ΕΔ

• ݸผϢʔβͷଐੑʹ͍ͭͯͷ৘ใ͸ಘΒΕͳ͍

• ͜ͷ৘ใΛ࢖ͬͯηάϝϯτͷਫ਼౓Λվળ͢Δ

View Slide

The PlaceIQ Analytic Platform: Location
Oriented Approaches to Mobile Audiences
• PlaceIQͱ͍͏Ґஔϕʔεͷ޿ࠂ෼ੳπʔϧͷ঺հ

View Slide

Practical Lessons from Predicting
Clicks on Ads at Facebook
• ஶ໊ͳΦϯϥΠϯ޿ࠂ࿦จ

• FBͷCTR༧ଌϞσϧͷ࣮૷ʹ͍ͭͯ

View Slide

iPinYou Global RTB Bidding
Algorithm Competition Dataset
• iPinYouͱ͍͏DSP͕ओ࠵ͨ͠RTBΞϧΰϦζϜͷίϯς
ετͰ࢖༻͞Εͨσʔληοτʹ͍ͭͯ

• ͜ͷσʔληοτ͸RTBܥϩδοΫͷ࣮ݧʹΑ͘࢖ΘΕͯ
͍Δ

View Slide

Can Television Advertising Impact Be Measured
on the Web? Web Spike Response as a Possible
Conversion Tracking System for Television.
• ςϨϏࢹௌऀ͸ಉ࣌ʹΠϯλʔωοτ΋΍ΔͷͰɺςϨϏ
CMͷࢹௌ͕े਺ඵޙͷωοτ׆ಈʹӨڹΛ༩͑Δ

• ͜ͷΑ͏ͳӨڹΛܭଌ͢ΔͨΊͷwebϕʔεγεςϜΛ։
ൃͨ͠

View Slide

TargetAd 2016
https://sites.google.com/site/targetad2016/program

View Slide

Distributed Representations of Web
Browsing Sequences for Ad Targeting
• Ϣʔβͷϒϥ΢δϯάཤྺΛಛ௃ྔԽ͍ͨ͠

• ϢʔβΛparagraphɺURLΛwordʹݟཱͯͯParagraph
Vector(W2Vͷ೿ੜɺύϥάϥϑͷembeddingΛಘΒΕΔ)
Λߏங͢Δͱ͍͏ͷΛલճ΍ͬͨ

• Backward PV-DMͱ͍͏վྑ൛ͷఏҊ

View Slide

Sponsored Video Advertising: Filling the
Void between Online and TV Advertising
• Ϣʔβ͕ಈըΛࢹௌ͍ͯ͠Δͱ͖ʹ޿ࠂΛݟ͍ͤͨɺݟͤ
·͘Γ͍ͨͱ͍͏ڧ͍ؾ͕࣋ͪ͋Δ

• ಈըʹؔ࿈͢Δ޿ࠂͷ՝ۚମܥʹ͍ͭͯͷఏҊ

View Slide

On Spatial-Aware Viral Marketing
for Location-based Advertisements,
• ৽͍͠ళΛ։͖͍ͨ৔߹ɺళͷ৔ॴɺޱίϛͷޮՌɺ෺
ཧ޿ࠂͷ৔ॴͳͲͷཁૉ͕ച্͛ʹӨڹΛ༩͑Δɻ

• ͦͷΑ͏ͳ੍໿ͷԼͰ࠷దͳީิΛબͿํ๏ΛఏҊ

View Slide

Keynote: Come back soon!
Estimating return times for users
• ߦಈཤྺΛݩʹɺϢʔβ͕࣍ʹ͜ͷαʔϏεΛ࢖͏ͷ͸
͍͔ͭਪఆ͍ͨ͠

• ੜଘ෼ੳɺ఺աఔɺLSTM

View Slide

Building a Bi-directed Recommendation
System for Mobile Users and App-install
Ad Campaigns
• ϞόΠϧΞϓϦʹؔ͢Δ૒ํ޲Ϩίϝϯσʔγϣϯγες
Ϝ

• ΞϓϦͷ޿ࠂΩϟϯϖʔϯʹରͯ͠ɺޮՌతͳϢʔβηά
ϝϯτΛਪન

• Ϣʔβʹରͯ͠ɺڵຯΛ࣋ͪͦ͏ͳΞϓϦΛਪન

View Slide

Finding Needle in a Million Metrics:
Anomaly Detection in a Large-scale
Computational Advertising Platform
• ޿ࠂϓϥοτϑΥʔϜʹ͓͚Δҟৗݕ஌

• Ξυϑϥ΢υͷจ຺Ͱ͸ͳ͘ɺγεςϜͷҟৗΛݕग़͢Δ
ͷ͕໨త

• ҆ఆͨ͠ύϑΥʔϚϯεΛग़͍ͯ͠ΔΩϟϯϖʔϯΛ؍ଌ

• ύϑΥʔϚϯεʹҟৗ͕͋ͬͨΒɺγεςϜଆʹݪҼ͕
͋ΔՄೳੑ͕ߴ͍

View Slide

Preserving Privacy in Geo-
Targeted Advertising
• ஍ҬͱऩೖΛ࢖ͬͨλʔήςΟϯά޿ࠂΛߟ͑Δ

• ߈ܸऀ͸τϐοΫϞσϦϯά΍ػցֶशΛ࢖͏͜ͱͰɺ͋
Δ޿ࠂ͕දࣔ͞ΕͨϢʔβͷ஍Ҭ΍ऩೖΛਪఆͰ͖Δ

• ഑৴͞ΕΔ޿ࠂʹϊΠζΛࠞͥΔ͜ͱͰ͜ͷΑ͏ͳਪఆΛ
๦͛Δ

View Slide

Invited talk: User
Recommendation in Instagram
• Instagramʹ͓͚Δ͓͢͢ΊϢʔβͷ࢓૊Έ

• ιʔγϟϧάϥϑͷߏ଄΍ڵຯͷ͋ΔτϐοΫΛ࢖ͬͨਪ
ન

View Slide

Demographic Prediction of Web
Requests from Labeled Aggregate Data
• Ϣʔβू߹ʹରͯ͠ɺσϞάϥଐੑͷׂ߹Λ஌Δ͜ͱ͕Ͱ
͖Δͱ͢Δ

• ͜ͷ৘ใΛ࢖ͬͯϢʔβ୯ҐͰͷଐੑਪఆΛ͢Δ

• Williams+, "Pleasing the advertising oracle: Probabilistic
prediction from sampled, aggregated ground truth",
AdKDD2014 ͷൃల

View Slide

Toward Personalized Product Search for
eCommerce Sites: A Case Study in Yahoo!
Taiwan
• Y!୆࿷ͷeίϚʔεαΠτʹ͓͚Δݕࡧ

• ଟ͘ͷݕࡧΫΤϦ͸ҰൠతͩͬͨΓᐆດͰɺϢʔβ͸๲େ
ͳ݁Ռͷத͔Β໨తͷ΋ͷΛ୳͞Ͷ͹ͳΒͳ͍

• ςΩετ৘ใ͚ͩͰͷϚονϯάͰ͸ෆे෼ͳͷͰɺϢʔ
βͷڵຯʹԠͯ͡ύʔιφϥΠζͨ͠ϥϯΩϯάΛߦ͏

View Slide

AdKDD & TargetAd
2017

View Slide

AdKDD & TargetAd 2017
https://adkdd-targetad.wixsite.com/adkddtargetad2017/accepted-papers

View Slide

Attribution Modeling Increases Efﬁciency of Bidding in
Display Advertising
• Criteo

• ΞτϦϏϡʔγϣϯϞσϧΛ࢖ͬͨbidઓུ

• 2nd price auctionʹ͓͚Δ࠷దઓུ͸ͦͷϦΫΤετͷՁ
஋Λೖࡳֹͱ͢Δ͜ͱ͕ͩɺਖ਼֬ͳՁ஋ΛٻΊΔͨΊʹ
͸ΞτϦϏϡʔγϣϯϞσϧΛߟྀ͢Δඞཁ͕͋Δ

View Slide

Blacklisting the Blacklist in
Online Advertising
• Dstillery

• RTBʹ͓͍ͯ͸publisherଆ͕ϒϥοΫϦετΛ͍࣋ͬͯͯ
ಛఆͷ޿ࠂΛ͸͘͜͡ͱ͕͋Δ

• ϒϥοΫϦετͷ಺༰͸ඇެ։

• ແବͳbidΛආ͚ΔͨΊϒϥοΫϦετͷ಺༰Λਪఆ͢Δ
ػೳΛ࡞ͬͨ

• γεςϜෛՙ͕ݮΓɺউ཰͕޲্ͨ͠

View Slide

Anti-Ad Blocking Strategy:
Measuring its True Impact
• Adobe Research

• AdBlockΛݕग़ͯ͠ϢʔβʹϝοηʔδΛग़͢ઓུ͕͋Δ
(՝ۚίʔε΁ͷ༠ಋͳͲ)

• ͜ͷઓུ͸Ͳͷఔ౓༗ޮͳͷ͔ௐ΂ΔͨΊͷ౷ܭతख๏
ΛఏҊ

• ͳΜΒ͔ͷࣄ৘ʹΑΓɺී௨ʹA/Bςετ͢ΔΘ͚ʹ͸͍
͔ͳ͍Β͍͠

View Slide

MM2RTB: Bringing Multimedia
Metrics to Real-Time Bidding
• ޿ࠂͷ࣭ΛධՁ͢ΔͨΊͷmultimedia metricsͱ͍͏ͷ͕
͋Δ

• contextual relevance, visual saliency, ad memorability

• ϖʔδ಺ʹෳ਺ͷ޿ࠂεϩοτ͕͋Δͱ͖ɺmultimedia
metricsΛߟྀͯ͠޿ࠂͷ૊Έ߹ΘͤΛબ୒͢Δख๏ͷఏ
Ҋ

View Slide

Data-Driven Reserve Prices for Social
Advertising Auctions at LinkedIn
• ηΧϯυϓϥΠεΦʔΫγϣϯʹ͓͚Δ࠷దͳϦβʔϒϓ
ϥΠεΛٻΊΔํ๏Λ2छྨ(Ϣʔβ/ηάϝϯτϨϕϧ)ఏ
Ҋ

• ࣮γεςϜʹ࣮૷ͯ͠ޮՌΛݕূͨ͠

View Slide

Optimal Reserve Price for Online Ads
Trading Based on Inventory Identiﬁcation
• Alibaba+Yahoo

• ͜Ε΋reserve priceͷܾఆઓུʹؔ͢Δ࿩

View Slide

Ranking and Calibrating Click-Attributed
Purchases in Performance Display
Advertising
• click-attributionͰCV՝ۚϞσϧͷ޿ࠂ഑৴

• bidͷࡍʹ͸ҎԼͷखॱ͕ඞཁ

• CV͞Ε΍͍͢޿ࠂͷީิΛྻڍ(ΞτϦϏϡʔγϣϯͷ
੍໿ͰɺΫϦοΫ཰΋ߴ͍ඞཁ͕͋Δ)

• CV཰ʹԠͯ͡ೖࡳֹۚΛܾఆ(ਖ਼֬ͳCV཰ͷਪఆ͸ࠔ
೉)

• ϥϯΩϯά+ΩϟϦϒϨʔγϣϯʹΑͬͯ͜ͷ໰୊Λղ͘

View Slide

Cost-sensitive Learning for Utility
Optimization in Online Advertising
Auctions
• CTR/CVR༧ଌʹ͸ɺhighly non-uniform misprediction
cost͕ଘࡏ͢Δ

• UtilityͷΑ͏ͳࢦඪΛ࢖ͬͯੑೳධՁ͢Δख๏͕ఏҊ͞Ε
͍ͯΔ

• ֶश࣌ʹ΋UtilityΛߟྀͯ͠࠷దԽ͍ͨ͠

• log lossʹॏΈΛ͚ͭΔ͜ͱͰੑೳվળ͢Δख๏ͷఏҊ

View Slide

A Practical Framework of Conversion Rate
Prediction for Online Display Advertising
• CVR༧ଌʹؔ͢Δ༷ʑͳख๏ͷఏҊ

• Over prediction΁ͷରԠ

• delayed feedbackΛߟֶྀͨ͠श

• ΞτϦϏϡʔγϣϯϞσϧΛߟྀͨ͠ೖࡳֹิਖ਼

View Slide

An Ensemble-based Approach to Click-
Through Rate Prediction for Promoted
Listings at Etsy
• Etsyͷlisting adͰ࢖ΘΕ͍ͯΔCTRਪఆγεςϜͷ঺հ

View Slide

Proﬁt Maximization for Online
Advertising Demand-Side Platform
• CPC/CPAϞσϧʹ͓͍ͯɺDSPͷརӹΛ࠷େԽ͍ͨ͠

• ࠷దͳೖࡳઓུ(Ͳͷ޿ࠂΛ͍͘ΒͰೖࡳ͢Δ͔)ʹ͍ͭͯ
ߟ࡯

• Ωϟϯϖʔϯ༧ࢉͱϖʔγϯάɺλʔήςΟϯάɺimpڙ
څྔͷ੍໿͕͋Δ

• ࣮༻తʹղ͚ΔϞσϧΛఏҊ

View Slide

Deep & Cross Network for
Ad Click Predictions
• Google

• Deep & Cross networkͱ͍͏DLͷϞσϧΛఏҊ

• ަޓ࡞༻Λཅʹѻ͑Δ

• CTRਪఆʹར༻ͯ͠ߴੑೳͩͬͨ

View Slide


View Slide

https://adkdd-targetad.wixsite.com/2018/accepted-papers

View Slide

A Large Scale Benchmark
for Uplift Modeling
• Criteo

• Uplift modelingͷݕূʹ࢖͑Δେن໛σʔλΛެ։ͨ͠

• https://ailab.criteo.com/criteo-uplift-prediction-dataset/

View Slide

Deep Density Networks and
Uncertainty in Recommender Systems
• Taboola(޿ࠂϓϥοτϑΥʔϜΛ΍ͬͯΔձࣾ)

• Deep Density Networksͱ͍͏ϞσϧͷఏҊ

• content based+ڠௐϑΟϧλϦϯάͷϋΠϒϦου

• ༧ଌͷෆ࣮֬ੑΛදݱͰ͖Δ

• ε-greedyͱUCBΛ૊Έ߹ΘͤͨΑ͏ͳϩδοΫͰ
exploration/exploitation͢Δ

View Slide

Deep Neural Net with Attention for
Multi-channel Multi-touch Attribution
• DNNΛ࢖ͬͨMulti-touch attributionϞσϧͷఏҊ

View Slide

Deep Policy Optimization for E-commerce
Sponsored Search Ranking Strategy
• 2018͔ΒDNNΛ࢖ͬͨൃද͕໨ཱͪ·͢Ͷ

• εϙϯαʔυαʔνͷϥϯΩϯά໰୊Λਂ૚ڧԽֶशͰ
ղ͘

View Slide

Designing Experiments to Measure
Incrementality on Facebook
• Facebookʹ͓͚Δincrementalityܭଌख๏ͷղઆ

• upliftͱincrementalityͷҧ͍͕Ṗ

• ୯ͳΔA/BςετΑΓॊೈͳ࣮ݧ͕Մೳ

View Slide

Dynamic Hierarchical Empirical Bayes: A
Predictive Model Applied to Online
Advertising
• Adobe

• εϙϯαʔυαʔνʹ͓͚ΔCTR/CVRਪఆϞσϧ

• ֊૚ϕΠζʹ͓͍ͯɺσʔλʹԠͯ͡ಈతʹ֊૚ߏ଄Λܾ
ఆ͢ΔΑ͏ͳϞσϧͷఏҊ

View Slide

Forecasting Granular Audience
Size for Online Advertising
• ಛఆͷηάϝϯτʹ޿ࠂ഑৴͢Δͱ͖ɺࣄલʹΦʔσΟΤ
ϯεͷαΠζ͕஌Γ͍ͨ

• Frequent Itemset Miningͱ͍͏ख๏ΛݩʹɺͦͷΑ͏ͳਪ
ఆΛ͢Δख๏ΛఏҊ

View Slide

Mini-Batch AUC
Optimization
• ֶश࣌ʹAUCΛ௚઀࠷దԽ͍͕ͨ͠ɺେن໛σʔληο
τʹ͓͍ͯ͸೉͔ͬͨ͠

• ϛχόονΛ࢖͏͜ͱͰߴ଎ͳֶशΛՄೳͱ͢Δख๏Λ
ఏҊ

View Slide

Optimal Bidding, Allocation and Budget
Spending for a Demand Side Platform
Under Many Auction Types
• ৽͍͠ೖࡳઓུϩδοΫͷఏҊ

• 1st/2nd price auctionʹରԠɺ༧ࢉ੍໿΍ΫϥΠΞϯτ͝
ͱͷر๬ΛߟྀͰ͖Δ

View Slide

AdKDD 2019

View Slide

AdKDD 2019
https://www.adkdd.org/2019-papers-and-talks

View Slide

Time-Aware Prospective Modeling of
Users for Online Display Advertising
• Yahoo

• ࣌ܥྻΛߟྀͨ͠ϢʔβߦಈϞσϦϯάʹΑΓɺϢʔβͷ
ҙਤΛߟྀͨ͠ਪఆΛߦ͏DNNϞσϧͷఏҊ

View Slide

Graphing Crumbling
Cookies
• ΫϩεσόΠετϥοΩϯάΛ͍͕ͨ͠ɺۙ೥cookieͷ࢖
༻੍ݶ͕ݫ͘͠ͳ͍ͬͯΔ(ITP౳)

• Browser ﬁngerprintingͳͲͷΘΔ͍ٕज़Λ࢖Θͣ͜ͷঢ়گ
ʹରԠ͢ΔͨΊͷख๏ΛఏҊ

View Slide

Causally Driven Incremental Multi Touch
Attribution Using a Recurrent Neural
Network
• JD.com(ژ౦঎৓): தࠃͷڊେECاۀ

• multi-touch attributionΛRNNͰϞσϦϯά͢Δख๏Λఏ
Ҋ

• ͜ͷϞσϧ͸࣮ࡍ࢖ΘΕ͍ͯΔΒ͍͠

View Slide

Combinatorial Keyword
Recommendations for Sponsored Search
with Deep Reinforcement Learning
• Baidu

• ਂ૚ڧԽֶशͰϨίϝϯμ࡞Δ΍ͭ

View Slide

Feasible Bidding Strategies
through Pure Exploration Bandits
• ೖࡳઓུͷީิ͕େྔʹ͋Γɺ͍͍΍ͭΛબͼ͍ͨ

• ύϥϝλͷ૊Έ߹ΘͤͰ͔ͳΓީิ͕ଟ͍(~100)ͨΊɺී
௨ʹόϯσΟοτΞϧΰϦζϜΛ࢖͏ͷ͸ඇޮ཰

• ʮͦͦ͜͜Αͦ͞͏ʯͳީิΛߴ଎ʹൃݟ͢ΔͨΊͷख๏
ΛఏҊ

View Slide

In-app Purchase Prediction Using
Bayesian Personalized DwellDay Ranking
• ژେ+CyberAgent

• ΞϓϦ಺՝ۚͦ͠͏ͳϢʔβΛਪఆ͢ΔλεΫ

• Bayesian personalized rankingΛݩʹͨ͠ख๏ΛఏҊ

View Slide

Learning from Multi-User Activity
Trails for B2B Ad Targeting
• Yahoo

• B2Bʹ͓͚Δ঎඼ͷߪೖϓϩηε͸ɺෳ਺ͷਓ͕ؒؔΘͬ
ͨෳࡶͳ෺ʹͳΔ

• ͜ͷΑ͏ͳϓϩηεʹରԠͨ͠޿ࠂͷ഑৴ઓུ͕ඞཁ

• ಉ͡૊৫ʹ͍ΔΦʔσΟΤϯεΛਪఆɺͦΕΒͷߦಈΛ૯
߹ͯ͠CV༧ଌ͢ΔϞσϧΛఏҊ

View Slide

Modeling Advertiser Bidding Behaviors in
Google Sponsored Search with a Mirror
Attention Mechanism
• SSPଆ͔ΒೖࡳऀͷߦಈΛϞσϦϯά͢Δ

• ৽ػೳΛϦϦʔεͨ͠ࡍɺͦΕ͕௕ظతʹ޿ࠂओʹͲͷΑ
͏ͳӨڹΛٴ΅͔͢஌Γ͍ͨ…… ͱ͍͏ͷ͕Ϟνϕʔ
γϣϯΒ͍͠

View Slide

Optimal bidding: a dual
approach
• ৽͍͠ೖࡳઓུͷఏҊ

View Slide

·ͱΊ
• ޿ࠂܥ࿦จ͕ͨ͘͞Μ͋Δ͜ͱ͕Θ͔Γ·ͨ͠

• ͍͔͕Ͱ͔ͨ͠?

• དྷ೥಄͘Β͍ʹ޿ࠂܥ࿦จಡॻձΛ΍Γ͍ͨͷͰɺڵຯ͋
Δਓ͸͓੠͕͚͍ͩ͘͞

View Slide

オンライン広告関連の論文を50本くらい雑に紹介する AdKDD編 / adkdd-all

オンライン広告関連の論文を50本くらい雑に紹介する AdKDD編 / adkdd-all

todesking

More Decks by todesking

Featured

Transcript