Deepayan Chakrabarti Erik Vee Traffic Shaping 2 Which article summary should be picked Ans The one with highest expected CTR Which ad should be displayed Ans The ad that minimizes ID: 1018553
Download Presentation The PPT/PDF document "1 Traffic Shaping to Optimize Ad Deliver..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. 1Traffic Shaping to Optimize Ad DeliveryDeepayan ChakrabartiErik Vee
2. Traffic Shaping2Which article summary should be picked?Ans: The one with highest expected CTRWhich ad should be displayed?Ans: The ad that minimizes underdeliveryArticle pool
3. UnderdeliveryAdvertisers are guaranteed some impressions (say, 1M) over some time (say, 2 months)only to users matching their specsonly when they visit certain types of pagesonly on certain positions on the pageAn underdelivering ad is one that is likely to miss its guarantee3
4. Traffic Shaping4Which article summary should be picked?Ans: The one with highest expected CTRWhich ad should be displayed?Ans: The ad that minimizes underdeliveryGoal: Combine the two
5. Traffic ShapingGoal: Bias the article summary selection toreduce under-deliverybut insignificant drop in CTRAND do this in real-time
6. OutlineFormulation as an optimization problemReal-time solutionEmpirical results6
7. Formulationj:(ads)ℓ:(user, article, position)“Fully Qualified Impression”i:(user, article)k:(user)ℓjikGoal: Infer traffic shaping fractions wkiSupply skCTR cki Traffic shaping fraction wkiDemand dj Ad delivery fraction φℓj
8. FormulationFull traffic shaping graph:All forecasted user traffic X all available articlesarriving at the homepage, or directly on article pageGoal: Infer wki But forced to infer φℓj as wellFull Traffic Shaping GraphABC Traffic shaping fraction wki Ad delivery fraction φℓjCTR cki
9. OutlineFormulation as an optimization problemReal-time solutionEmpirical results9
10. FormulationReformulation: {wki, φℓj}→ zℓjConvex program can be solved optimally10
11. FormulationBut we have another problemAt runtime, we must shape every incoming user without looking at the entire graphSolution:Periodically solve the convex problem offlineStore a cache derived from this solutionReconstruct the optimal solution for each user at runtime, using only the cache11
12. Real-time solution12Cache theseReconstruct using theseAll constraints can be expressed as constraints on σℓ
13. ResultsData: Historical traffic logs from April, 201125K user nodesTotal supply weight > 50B impressions100K ads13
14. Lift in impressionsLift in impressions delivered to underperforming adsFraction of traffic that is not shapedNearly threefold improvement via traffic shaping14
15. Average CTRAverage CTR (as percentage of maximum CTR)Fraction of traffic that is not shapedCTR drop < 10%15
16. ResultsData: Historical traffic logs from April, 201125K user nodesTotal supply weight > 50B impressions100K ads3x underdelivery reduction with <10% CTR drop16
17. Summary3x underdelivery reduction with <10% CTR drop2.6x reduction with 4% CTR dropRuntime application needs only a small cache17
18. UnderdeliveryHow can underdelivery be computed?Need user traffic forecastsDepends on other ads in the systemAn ad-serving systemwill try to minimizeunder-delivery on thisgraph18Forecasted impressions(user, article, position)Ad inventorySupply sℓDemand dj ℓj
19. Real-time solution1912σℓ = 0 unless Σzℓj = maxℓ Σzℓj3Σℓ σℓ = constant for all i connected to kΣzℓj UiLiσℓ 3 KKT conditionsShape depends on the cached duals αjℓjki
20. Real-time solution2012σℓ = 0 unless Σzℓj = maxℓ Σzℓj3Σℓ σℓ = constant for all i connected to kℓjkiΣzℓj UiLiσℓ Algo Initialize σℓ = 0 Compute Σzℓj from (1) If constraints unsatisfied, increase σℓ while satisfying (2) and (3) Repeat Extract wki from zℓj
21. Comparison with other methods21
22. Key TransformationThis allows a reformulation solely in terms of new variables zℓj zℓj = fraction of supply that is shown ad j, assuming user always clicks article22
23. ResultsData: Historical traffic logs from April, 201125K user nodesTotal supply weight > 50B impressions100K adsWe compare our model to a scheme thatpicks articles to maximize expected CTR, andpicks ads to display via a separate greedy method23
24. Formulation24ℓjikunderdeliveryTotal user traffic flowing to j (accounting for CTR loss)demand(Satisfy demand constraints)skwkicki
25. Formulation25ℓjik(Bounds on traffic shaping fractions)(Shape only available traffic)(Satisfy demand constraints)(Ad delivery fractions)