1Challenges in Computational AdvertisingDeepayan Chakrabarti (deepay@yahoo-inc.c

1Challenges in Computational AdvertisingDeepayan Chakrabarti (deepay@yahoo-inc.c www.phwiki.com

1Challenges in Computational AdvertisingDeepayan Chakrabarti (deepay@yahoo-inc.c

Barker, Jay, Morning Show Host has reference to this Academic Journal, PHwiki organized this Journal 1Challenges in Computational AdvertisingDeepayan Chakrabarti (deepay@yahoo-inc.com)Online Advertising OverviewAdvertisersAd NetworkAdsContentPick adsUserContent ProviderExamples: Yahoo, Google, MSN, RightMedia, 2Advertising SettingDisplayContent MatchSponsored Search

Sheffield Hallam University GB www.phwiki.com

This Particular University is Related to this Particular Journal

Advertising SettingPick adsDisplayContent MatchSponsored Search4Advertising SettingGraphical display adsMostly as long as br in addition to awarenessRevenue based on number of impressions (not clicks)DisplayContent MatchSponsored Search5Advertising SettingContent match adDisplayContent MatchSponsored Search6

Advertising SettingPick adsText adsMatch ads to the contentDisplayContent MatchSponsored Search7Advertising SettingThe user intent is unclearRevenue depends on number of clicksQuery (webpage) is long in addition to noisyDisplayContent MatchSponsored Search8Advertising SettingSearch QuerySponsored Search AdsDisplayContent MatchSponsored Search9

This presentationContent Match [KDD 2007]: How can we estimate the click-through rate (CTR) of an ad on a page~106 ads~109 pagesCTR as long as ad j on page i10This presentationEstimating CTR as long as Content Match [KDD ‘07]Traffic Shaping as long as Display Advertising [EC ‘12]Article summaryAlternatesclickDisplay ads11This presentationEstimating CTR as long as Content Match [KDD ‘07]Traffic Shaping as long as Display Advertising [EC ‘12]Recommend articles (not ads)need high CTR on article summaries+ prefer articles on which under-delivering ads can be shown12

This presentationEstimating CTR as long as Content Match [KDD ‘07]Traffic Shaping as long as Display Advertising [EC ‘12]Theoretical underpinnings [COLT ‘10 best student paper]Represent relationships as a graphRecommendation = Link PredictionMany useful heuristics existWhy do these heuristics work1314Estimating CTR as long as Content MatchContextual AdvertisingShow an ad on a webpage (“impression”)Revenue is generated if a user clicksProblem: Estimate the click-through rate (CTR) of an ad on a pageEstimating CTR as long as Content MatchWhy not use the MLEFew (page, ad) pairs have N>0Very few have c>0 as wellMLE does not differentiate between 0/10 in addition to 0/100We have additional in as long as mation: hierarchies15

16Estimating CTR as long as Content MatchUse an existing, well-understood hierarchyCategorize ads in addition to webpages to leaves of the hierarchyCTR estimates of siblings are correlatedThe hierarchy allows us to aggregate dataCoarser resolutionsprovide reliable estimates as long as rare eventswhich then influences estimation at finer resolutions17Estimating CTR as long as Content MatchLevel 0Level iPage hierarchyAd hierarchyRegion = (page node, ad node)Region HierarchyA cross-product of the page hierarchy in addition to the ad hierarchyPage classesAd classesRegionEstimating CTR as long as Content MatchOur ApproachData Trans as long as mationModelModel Fitting18

Data Trans as long as mationProblem:Solution: Freeman-Tukey trans as long as mDifferentiates regions with 0 clicksVariance stabilization:19ModelGoal: Smoothing across siblings in hierarchy [Huang+Cressie/2000]2020Level iLevel i+1S1S2S3S4Sparent Each region has a latent state Sr yr is independent of the hierarchy given Sr Sr is drawn from its parent Spa(r) Model21SrSpa(r)yrypa(r)variance Vrvariance wrVpa(r)wpa(r)

However, learning Wr , Vr in addition to r as long as each region is clearly infeasibleAssumptions:All regions at the same level share the same W() in addition to () Vr = V/Nr as long as some constant V, sinceModel22ModelImplications: determines degree of smoothing :Sr varies greatly from Spa(r) Each region learns its own Sr No smoothing :All Sr are identicalA regression model on features ur is learntMaximum Smoothing23Implications: determines degree of smoothingVar(Sr) increases from root to leafBetter estimates at coarser resolutionsModel24

Barker, Jay WJOX-FM Morning Show Host www.phwiki.com

Implications: determines degree of smoothingVar(Sr) increases from root to leafCorrelations among siblings at level :Depends only on level of least common ancestorModel25Estimating CTR as long as Content MatchOur ApproachData Trans as long as mation (Freeman-Tukey)Model (Tree-structured Markov Chain)Model Fitting2627Model FittingFitting using a Kalman filtering algorithmFiltering: Recursively aggregate data from leaves to rootSmoothing: Propagate in as long as mation from root to leavesComplexity: linear in the number of regions, as long as both time in addition to spacefilteringsmoothing

28Model FittingFitting using a Kalman filtering algorithmFiltering: Recursively aggregate data from leaves to rootSmoothing: Propagates in as long as mation from root to leavesKalman filter requires knowledge of , V, in addition to WEM wrapped around the Kalman filterfilteringsmoothing29Experiments503M impressions7-level hierarchy of which the top 3 levels were usedZero clicks in76% regions in level 295% regions in level 3Full dataset DFULL, in addition to a 2/3 sample DSAMPLE30ExperimentsEstimate CTRs as long as all regions R in level 3 with zero clicks in DSAMPLESome of these regions R>0 get clicks in DFULLA good model should predict higher CTRs as long as R>0 as against the other regions in R

ModelGoal: Smoothing across siblings in hierarchyOur approach:Each region has a latent state Sr yr is independent of hierarchy given Sr Sr is drawn from the parent region Spa(r) 7373Data Trans as long as mationProblem:Solution: Freeman-Tukey trans as long as mDifferentiates regions with 0 clicksVariance stabilization:74MLE CTRN Var(MLE)Mean yrN Var(yr)

Barker, Jay Morning Show Host

Barker, Jay is from United States and they belong to WJOX-FM and they are from  Birmingham, United States got related to this Particular Journal. and Barker, Jay deal with the subjects like Sports

Journal Ratings by Sheffield Hallam University

This Particular Journal got reviewed and rated by Sheffield Hallam University and short form of this particular Institution is GB and gave this Journal an Excellent Rating.