This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

Aranyak Mehta Vahab Mirrokni

June 7, 2011

Outline of this talk

**Ad delivery for contract-based settings
**

Ad Serving Planning

**Ad serving in repeated auction settings
**

General architecture. Allocation for budget constrained advertisers.

Other interactions

Learning + allocation Learning + auction Auction + contracts

Contract-based Ad Delivery: Outline

**Basic Information Ad Serving.
**

Targeting. Online Allocation

Ad Planning: Reservation

Contract-based Online Advertising Pageviews (impressions) instead of queries. Video Ads. Mobile Ads. . Display/Banner Ads.

Mobile Ads. Cost-Per-Impression (CPM). .Contract-based Online Advertising Pageviews (impressions) instead of queries. Display/Banner Ads. Video Ads. Not Auction-based: oﬄine negotiations + Online allocations.

7 billion. . Verizon. Not Auction-based: oﬄine negotiations + Online allocations. Cost-Per-Impression (CPM). Scottrade. Top Advertiser: AT&T. $2. Display/Banner Ads: Q1. Mobile Ads. Display/Banner Ads.Contract-based Online Advertising Pageviews (impressions) instead of queries. 2010: One Trillion Display Ads in US. Video Ads.

$2. .Contract-based Online Advertising Pageviews (impressions) instead of queries. Facebook. 2010: One Trillion Display Ads in US. Mobile Ads. Scottrade. Ad Serving Systems e. Display/Banner Ads. Video Ads.g. Google Doubleclick. Cost-Per-Impression (CPM).7 billion. Not Auction-based: oﬄine negotiations + Online allocations. Display/Banner Ads: Q1. Verizon. Top Advertiser: AT&T. AdMob.

**Internet Advertising Revenues - 2010
**

Other 6% Classified 10% Search 46%

Internet Ad Revenues - 2010

Display 38%

24% increase

Total $26.0 billion

**Display Ad Delivery: Overview
**

Display Ad Delivery Planning:

Ad Serving: Targeting:

Allocation:

**1. Planning: Contracts/Commitments with Advertisers. 2. Ad Serving:
**

Targeting: Predicting value of impressions. Ad Allocation: Assigning Impressions to Ads Online.

**Display Ad Delivery: Overview
**

Display Ad Delivery Planning:

Oﬄine, Online Strategic, Stochastic

Ad Serving: Targeting:

Allocation:

Online, Stochastic

**1. Planning: Contracts/Commitments with Advertisers. 2. Ad Serving:
**

Targeting: Predicting value of impressions. Ad Allocation: Assigning Impressions to Ads Online.

Display Ad Delivery: Overview Display Ad Delivery Planning: Oﬄine. Planning: Contracts/Commitments with Advertisers. Ad Allocation: Assigning Impressions to Ads Online. 2. Online Strategic. Ad Serving: Targeting: Predicting value of impressions. . Stochastic Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: Allocation: Online. Stochastic 1.

Ad Serving: Targeting: Predicting value of impressions. Online Strategic. Budget Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: Allocation: Online. Stochastic Delivery Constraints. Stochastic 1.Display Ad Delivery: Overview Display Ad Delivery Planning: Oﬄine. . 2. Planning: Contracts/Commitments with Advertisers. Ad Allocation: Assigning Impressions to Ads Online.

Planning: Contracts/Commitments with Advertisers. Ad Allocation: Assigning Impressions to Ads Online.Display Ad Delivery: Overview Display Ad Delivery Planning: Oﬄine. Ad Serving: Targeting: Predicting value of impressions. Stochastic 1. . Online Strategic. Stochastic Delivery Constraints. 2. Budget Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: CTR Allocation: Online.

Smoothness. Budget Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: CTR Allocation: Online.Display Ad Delivery: Overview Display Ad Delivery Planning: Oﬄine. Fairness. Stochastic Objective Functions: Eﬃciency: Users and Advertisers. Stochastic Delivery Constraints. Revenue of the Publisher. . Delivery Penalty. Online Strategic.

Targeting.Contract-based Ad Delivery: Outline Basic Information Ad Serving. Online Ad Allocation Ad Planning: Reservation .

.Targeting Estimating Value of an impression.

Wang. Jiang. Behavioral Targeting Interest-based Advertising. 2009. Yan. Chen. Zhang.Targeting Estimating Value of an impression. How much can Behavioral Targeting Help Online Advertising? . Liu.

Targeting Estimating Value of an impression. Behavioral Targeting Interest-based Advertising. Yan. A semantic approach to contextual advertising . Riedel. Chen. How much can Behavioral Targeting Help Online Advertising? Contextual Targeting Information Retrieval (IR). Fontoura. Broder. Wang. Josifovski. Liu. Jiang. Zhang. 2009.

Liu. How much can Behavioral Targeting Help Online Advertising? Contextual Targeting Information Retrieval (IR).Targeting Estimating Value of an impression. Zhang. Chen. Wang. Riedel. Broder. Josifovski. 2009. Fontoura. Yan. Behavioral Targeting Interest-based Advertising. A semantic approach to contextual advertising Creative Optimization Experimentation . Jiang.

Predicting value of Impressions for Display Ads Estimating Click-Through-Rate (CTR). Budgeted Multi-armed Bandit Probability of Conversion. .

Short-term value of display ads? Archak. Budgeted Multi-armed Bandit Probability of Conversion. Computing Adfactors based on AdGraphs Markov Models for Advertiser-speciﬁc User Behavior . Mirrokni. 2010 Graph-based Models.Predicting value of Impressions for Display Ads Estimating Click-Through-Rate (CTR). Long-term vs. Muthukrishnan.

Online Ad Allocation . Targeting.Contract-based Ad Delivery: Outline Basic Information Ad Planning: Reservation Ad Serving.

Outline: Online Allocation Online Stochastic Assignment Problems Online (Stochastic) Matching Online Stochastic Packing Online Generalized Assignment (with free disposal) Experimental Results Online Learning and Allocation .

assign an eligible ad. value of assigning page i to ad a: via .Online Ad Allocation When page arrives.

value of assigning page i to ad a: via Display Ads (DA) problem: Maximize value of ads served: max Capacity of ad a: i∈A(a) xia ≤ Ca i.a via xia .Online Ad Allocation When page arrives. assign an eligible ad.

Online Ad Allocation When page arrives. revenue from assigning page i to ad a: bia “AdWords” (AW) problem: Maximize revenue of ads served: max Budget of ad a: i∈A(a) bia xia ≤ Ba i.a bia xia . assign an eligible ad.

General Form of LP max i. Ads (DA): sia = 1 AdWords (AW): sia = via xia ≤ 1 a sia xia ≤ Ca i xia ≥ 0 Online Matching: via = sia = 1 .a via xia (∀ i) (∀ a) (∀ i. a) Disp.

a via xia (∀ i) (∀ a) (∀ i. 1 [KVV]: 1− e -aprx Worst-Case .General Form of LP max i.BJN]: 1 1− e -aprx xia ≤ 1 a sia xia ≤ Ca i xia ≥ 0 Online Matching: via = sia = 1 1 Greedy: 2 . a) Disp. Ads (DA): sia = 1 AdWords (AW): sia = via [MSVV.

Ads (DA): sia = 1 ? AdWords (AW): sia = via [MSVV. 2 1 [KVV]: 1− e -aprx Disp.Ad Allocation: Problems and Models Online Matching: via = sia = 1 Greedy: 1 .BJN]: 1 1− e -aprx Worst Case .

i.) ? Stochastic i. 2 1 [KVV]: 1− e -aprx Disp.d model with unknown distribution) . Ads (DA): sia = 1 ? AdWords (AW): sia = via [MSVV.BJN]: 1 1− e -aprx [DH09]: 1− -aprx.i.d model with known distribution random order model (i.i.Ad Allocation: Problems and Models Online Matching: via = sia = 1 Greedy: 1 .i. if opt max via Worst Case Stochastic (i.d model: i.d.

702-aprx ? i.d with known distribution AdWords (AW): sia = via [MSVV.MOS11]: 0. if opt max via Worst Case Stochastic (i.BJN]: 1 1− e -aprx [DH09]: 1− -aprx.i.i.) Stochastic i. ? 1 [KVV]: 1− e -aprx [FMMM09.i. Ads (DA): via = sia = 1 sia = 1 1 Greedy: 2 .d.i.d model with unknown distribution) .d model: i.i.Ad Allocation: Problems and Models Online Matching: Disp.d model with known distribution random order model (i.

i. Ads (DA): sia = 1 ? ? AdWords (AW): sia = via [MSVV. 2 1 [KVV]: 1− e -aprx [FMMM09]: 0. if opt max via .Ad Allocation: Problems and Models Worst Case Stochastic (i.d with known distribution Disp.67-aprx i.i.BJN]: 1 1− e -aprx [DH09]: 1− -aprx.d.) Online Matching: via = sia = 1 Greedy: 1 .

i.AWY]: 1− -aprx.i.d.MOS11]: [FHKMS10. 0.d.) Online Matching: Disp.BJN]: 1 1− e -aprx [DH09]: 1− -aprx. if opt max via random order = i.i.702-aprx max via i.d with known if opt and Ca max sia distribution AdWords (AW): sia = via [MSVV. 2 ? 1 [KVV]: 1− e -aprx [FMMM09. Ads (DA): via = sia = 1 sia = 1 Greedy: 1 . model with unknown distribution .Ad Allocation: Problems and Models Worst Case Stochastic (i.

a) βa . zi ≥ 0 (∀i.Stochastic DA: Dual Algorithm max i. a) xia ≤ Ca i xia ≥ 0 .a via xia (∀ i) (∀ a) (∀ i. a) min a Ca βa + i zi xia ≤ 1 a zi ≥ via − βa (∀i.

zi ≥ 0 (∀i. Learn a dual variable for each ad βa . a) min a Ca βa + i zi xia ≤ 1 a zi ≥ via − βa (∀i.a via xia (∀ i) (∀ a) (∀ i. a) xia ≤ Ca i xia ≥ 0 Algorithm: Observe the ﬁrst fraction sample of impressions. . by solving the dual program on the sample. a) βa .Stochastic DA: Dual Algorithm max i. Assign each impression i to ad a that maximizes via − βa .

this algorithm is a (1−O( ))-aprx.AWY]: W. Henzinger.. Korula.h. as long as each item has low value (via ≤ moptn ). and large log capacity (Ca ≥ m log n 3 ) . Stein 2010 Thm[FHKMS10.Stochastic DA: Dual Algorithm Feldman.p. M.

this algorithm is a (1−O( ))-aprx. slackness. Given βa . as long as each item has low value (via ≤ moptn ).p. compute x ∗ as follows: ∗ ∗ xia = 1 if a = argmax(via − βa ). M.. ﬁnds opt ∗ Proof: Comp. and large log capacity (Ca ≥ m log n 3 ) ∗ Fact: If optimum βa are known. Henzinger. .Stochastic DA: Dual Algorithm Feldman.h. this alg.AWY]: W. Korula. Stein 2010 Thm[FHKMS10.

. M. Henzinger. the sample βa are ∗ close to βa . W. Stein 2010 Thm[FHKMS10.. . this algorithm is a (1−O( ))-aprx.p.AWY]: W.h. this alg. as long as each item has low value (via ≤ moptn ). Korula. ﬁnds opt ∗ Proof: Comp. compute x ∗ as follows: ∗ ∗ xia = 1 if a = argmax(via − βa ). Given βa .Stochastic DA: Dual Algorithm Feldman. Lemma: In the random order model.h. Extending DH09. slackness. and large log capacity (Ca ≥ m log n 3 ) ∗ Fact: If optimum βa are known.p.

Choose o ∈ Oi . the PD algorithm is a (1−O( ))-aprx.AWY]: W.General Stochastic Packing LPs m ﬁxed resources with capacity Ca Items i arrive online with options Oi . 3C a opt log n ) and . Thm[FHKMS10.p.h. values vio . as long as items have low value (vio ≤ small size (sioa ≤ log n ). using up capacity sioa in all a. use sioa . rsrc.

Wang. as long as items have low value (vio ≤ small size (sioa ≤ log n ).h. Thm[FHKMS10. Choose o ∈ Oi . use sioa . 3C a opt log n ) and Other Results and Extensions (random order model): Agrawal.p. Ye: Updating dual variables by periodic 2C solution of the dual program: Ca ≥ m log n or sioa ≤ M a 2 . using up capacity sioa in all a.General Stochastic Packing LPs m ﬁxed resources with capacity Ca Items i arrive online with options Oi . rsrc. values vio .AWY]: W. the PD algorithm is a (1−O( ))-aprx.

AWY]: W. Ye: Updating dual variables by periodic 2C solution of the dual program: Ca ≥ m log n or sioa ≤ M a 2 Vee. Thm[FHKMS10. Vassilvitskii . 3C a opt log n ) and Other Results and Extensions (random order model): Agrawal. Wang. as long as items have low value (vio ≤ small size (sioa ≤ log n ). values vio . using up capacity sioa in all a. Shanmugasundaram 2010: extension to convex objective functions: Using KKT conditions.p.General Stochastic Packing LPs m ﬁxed resources with capacity Ca Items i arrive online with options Oi .h. . rsrc. Choose o ∈ Oi . use sioa . the PD algorithm is a (1−O( ))-aprx.

Ads (DA): via = sia = 1 sia = 1 Greedy: 1 .MOS11]: [FHKMS10.BJN]: 1 1− e -aprx [DH09]: 1− -aprx. 2 ? 1 [KVV]: 1− e -aprx [FMMM09.702-aprx max via i.Ad Allocation: Problems and Models Worst Case Stochastic (i.AWY]: 1− -aprx.) Online Matching: Disp.i. if opt max via .i. 0.d with known if opt and Ca max sia distribution AdWords (AW): sia = via [MSVV.d.

d.) Free Disposal [FKMMP09]: 1 1− e -aprx: if Ca max sia [FMMM09. 0.d with known if opt and Ca max sia distribution . if opt max via Stochastic (i.MOS11]: [FHKMS10. 2 1 [KVV]: 1− e -aprx Disp.AWY]: 1− -aprx.i.Ad Allocation: Problems and Models Online Matching: via = sia = 1 Worst Case Greedy: 1 .i.702-aprx max via i.BJN]: 1 1− e -aprx [DH09]: 1− -aprx. Ads (DA): sia = 1 AdWords (AW): sia = via [MSVV.

07 Ad 1: C1 = 1 .DA: Free Disposal Model 0.

DA: Free Disposal Model 0.07 Ad 1: C1 = 1 .

DA: Free Disposal Model 0.07 Ad 1: C1 = 1 0.7 .

but no bonus points for extra impressions.07 Ad 1: C1 = 1 0. .7 Advertisers may not complain about extra impressions. either.DA: Free Disposal Model 0.

. Value of advertiser = sum of values of top Ca items she gets.7 Advertisers may not complain about extra impressions. but no bonus points for extra impressions.DA: Free Disposal Model 0. either.07 Ad 1: C1 = 1 0.

impression value).min. .Greedy Algorithm Assign impression to an advertiser maximizing Marginal Gain = (imp. value .

Greedy Algorithm Assign impression to an advertiser maximizing Marginal Gain = (imp. Competitive Ratio: 1/2. impression value).min. . [NWF78] Follows from submodularity of the value function. value .

Competitive Ratio: 1/2.Greedy Algorithm Assign impression to an advertiser maximizing Marginal Gain = (imp. impression value). 1 1+ n copies Ad 1: C1 = n Ad 2: C2 = n .min. value . [NWF78] Follows from submodularity of the value function.

1 1+ n copies Ad 1: C1 = n n copies 1 Ad 2: C2 = n . [NWF78] Follows from submodularity of the value function. impression value). Competitive Ratio: 1/2.Greedy Algorithm Assign impression to an advertiser maximizing Marginal Gain = (imp. value .min.

Competitive Ratio: 1/2. value . 1 1+ n copies Ad 1: C1 = n n copies 1 Ad 2: C2 = n Evenly Split? .Greedy Algorithm Assign impression to an advertiser maximizing Marginal Gain = (imp. impression value).min. [NWF78] Follows from submodularity of the value function.

A better algorithm? Assign impression to an advertiser a maximizing (imp. value . where βa = average value of top Ca impressions assigned to a.βa ). .

where βa = average value of top Ca impressions assigned to a.A better algorithm? Assign impression to an advertiser a maximizing (imp.βa ). n copies 1 1+ Ad 1: C1 = n n copies 1 Ad 2: C2 = n . value .

[FKMMP09] Primal-Dual Approach.A better algorithm? Assign impression to an advertiser a maximizing (imp. n copies 1 1+ Ad 1: C1 = n n copies 1 2 1 Ad 2: C2 = n Competitive Ratio: if Ca >> 1. . value .βa ). where βa = average value of top Ca impressions assigned to a.

. Better (pd-avg): βa = average value of top Ca impressions assigned to a.An Optimal Algorithm Assign impression to an advertiser a: maximizing (imp. Greedy: βa = min.βa ). value . impression assigned to a.

≥ v (Ca ): βa = 1 Ca (e − 1) Ca v (j)(1 + j=1 1 j−1 ) . Greedy: βa = min. .An Optimal Algorithm Assign impression to an advertiser a: maximizing (imp. value . Ca . Optimal (pd-exp): order value of edges assigned to a: v (1) ≥ v (2) . Better (pd-avg): βa = average value of top Ca impressions assigned to a.βa ). . impression assigned to a.

An Optimal Algorithm Assign impression to an advertiser a: maximizing (imp. Korula. [Feldman.βa ). Ca 1 Thm: pd-exp achieves optimal competitive Ratio: 1 − e − if Ca > O( 1 ). impression assigned to a. value . Better (pd-avg): βa = average value of top Ca impressions assigned to a.. Greedy: βa = min. . Pal 2009] . ≥ v (Ca ): βa = 1 Ca (e − 1) Ca v (j)(1 + j=1 1 j−1 ) . Muthukrishnan. . Optimal (pd-exp): order value of edges assigned to a: v (1) ≥ v (2) . M.

DA: sia = 1. zi ≥ 0 sia xia ≤ Ca i xia ≥ 0 . a) min a Ca β a + i zi (∀i. AW: via = sia .a via xia (∀ i) (∀ a) (∀ i. a) xia ≤ 1 a sia βa + zi ≥ via βa . max i. a) (∀i.Online Generalized Assignment (with free disposal) Multiple Knapsack: Item i may have diﬀerent value (via ) and diﬀerent size sia for diﬀerent ads a.

a) xia ≤ 1 a sia βa + zi ≥ via βa . zi ≥ 0 sia xia ≤ Ca i xia ≥ 0 Oﬄine Optimization: 1 − − δ-aprx[FGMS07.FV08].Online Generalized Assignment (with free disposal) Multiple Knapsack: Item i may have diﬀerent value (via ) and diﬀerent size sia for diﬀerent ads a. a) (∀i. max i. 1 e Thm[FKMMP09]: There exists a 1 − C algorithm if maxasia ≥ 1 . − -approximation . DA: sia = 1.a via xia (∀ i) (∀ a) (∀ i. AW: via = sia . a) 1 e min a Ca β a + i zi (∀i.

a) (∀i.Proof Idea: Primal-Dual Analysis [BJN] max i.a via xia (∀ i) min (∀ a) (∀ i. a) a xia ≤ 1 a Ca βa + i zi (∀i. zi ≥ 0 . a) sia xia ≤ Ca i sia βa + zi ≥ via xia ≥ 0 βa .

a) a xia ≤ 1 a Ca βa + i zi (∀i. update x. and zi = 0.. 2. Start from feasible primal and dual (xia = 0.e. . zi ≥ 0 1. 1 3. Primal=Dual=0). a) sia xia ≤ Ca i sia βa + zi ≥ via xia ≥ 0 Proof: βa . βa = 0.a via xia (∀ i) min (∀ a) (∀ i. i. β. z variables and keep primal and dual solutions. After each assignment. Show ∆(Dual) ≤ (1 − e )∆(Primal).Proof Idea: Primal-Dual Analysis [BJN] max i. a) (∀i.

Ad Allocation: Problems and Models Online Matching: via = sia = 1 Worst Case Greedy: 1 .AWY]: [FMMM09.MOS11]:1− -aprx. 0. if opt max via Stochastic (random arrival order) Free Disposal [FKMMP09]: 1 1− e -aprx: if Ca max sia [FHKMS10.BJN]: 1 1− e -aprx [DH09]: 1− -aprx. Ads (DA): sia = 1 AdWords (AW): sia = via [MSVV.702-aprx if opt max via and Ca max sia . 2 1 [KVV]: 1− e -aprx Disp.

Outline: Online Allocation Online Stochastic Assignment Problems Online Stochastic Packing Online Generalized Assignment (with free disposal) Experimental Evaluation Online Stochastic Weighted Matching .

.Dual-based Algorithms in Practice Algorithm: Assign each item i to ad a that maximizes via − βa .

More practical compared to Primal Algorithms: Just keep one number βa per advertiser. . Suitable for Distributed Ad Serving Schemes.Dual-based Algorithms in Practice Algorithm: Assign each item i to ad a that maximizes via − βa .

Dual-based Algorithms in Practice Algorithm: Assign each item i to ad a that maximizes via − βa . More practical compared to Primal Algorithms: Just keep one number βa per advertiser. Training-based Algorithms Compute βa based on historical/sample data. . Suitable for Distributed Ad Serving Schemes.

Training-based Algorithms Compute βa based on historical/sample data.Dual-based Algorithms in Practice Algorithm: Assign each item i to ad a that maximizes via − βa . blend in online algorithm. . More practical compared to Primal Algorithms: Just keep one number βa per advertiser. Suitable for Distributed Ad Serving Schemes. Hybrid approach (see also [MNS07]): Start with trained βa (past history).

5M impressions in simulation period 100 . dualbase: training-based primal-dual [FHKMS10] hybrid: convex combo of training based.2600 advertisers Edge weights = predicted click probability Eﬃciency: free disposal model Algorithms: greedy: maximum marginal value pd-avg.1. pure online. pd-exp: pure online primal-dual from [FKMMP09].Experiments: setup Real ad impression data from several large publishers 200k . lp-weight: optimum eﬃciency .

) dualbase outperforms pure online algorithms by 6% to 12%. Hybrid has a mild improvement of 2% (up to 10%). . pd-avg performs much better than the theoretical analysis.Experimental Evaluation: Summary Algorithm opt greedy pd-avg pd-exp dualbase hybrid Avg Eﬃciency% 100 69 77 82 87 89 pd-exp & pd-avg outperform greedy by 9% and 14% (with more improvements in tight competition.

” .Other Metrics: Fairness Qualititative deﬁnition: advertisers are “treated equally.

measure 1 distance to x ∗ .” One suggestion[FHKMS10]: Compute ”fair” solution x ∗ . .Other Metrics: Fairness Qualititative deﬁnition: advertisers are “treated equally.

Fair solution: Each a chooses best Ca impressions (highest via ) Repeat: Impressions shared among those who chose them. measure 1 distance to x ∗ .” One suggestion[FHKMS10]: Compute ”fair” solution x ∗ . a chooses an additional imp.Other Metrics: Fairness Qualititative deﬁnition: advertisers are “treated equally. If some a not receiving Ca imps. .

0 0 0.Experiments: highlights 1.0 q lp_weight 40 lp−weight fair dualbase pd−avg 50 Fairness pd−exp q 30 Efficiency (relative) 100 greedy q 20 q hybriddualbase pd−avg q q 10 fair q 40 50 60 70 80 90 0.6 0.4 0.8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Efficiency Advertiser .2 0.

2 0.4 0.Experiments: highlights 1.0 q dualbase Efficiency (relative) 15 pd−exp q q pd−avg Fairness 10 5 70 75 80 85 90 95 100 0.6 0.0 q fair 0 0.8 q greedy hybrid q q lp_weight lp−weight fair dualbase pd−avg 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Efficiency Advertiser .

M. 2011 .In Production Smooth Delivery of Display Ads Delivery of impressions throughout time should follow the traﬃc smoothly.. Feldman. Bhalgat. Model this with multiple nested capacity constraints: 1 − 1/e-competitive algorithm for this extension.

Feldman.In Production Smooth Delivery of Display Ads Delivery of impressions throughout time should follow the traﬃc smoothly... Model this with multiple nested capacity constraints: 1 − 1/e-competitive algorithm for this extension. Muthukrishnan. 2011 . M. Feldman. Bhalgat. 2011 Combined Allocation with Ad Exchange Yield Optimization of Display Advertising with Ad Exchange Belsairo. M.

2011 Re-act adaptively and quickly to changes in traﬃc: Use a control loop on the dual variable... Muthukrishnan. Tan and R. Bhalgat. Feldman. Srikant. 2011 . M. M. 2011 Combined Allocation with Ad Exchange Yield Optimization of Display Advertising with Ad Exchange Belsairo.In Production Smooth Delivery of Display Ads Delivery of impressions throughout time should follow the traﬃc smoothly. Assymptotically optimal policy: B. Model this with multiple nested capacity constraints: 1 − 1/e-competitive algorithm for this extension. Feldman.

Online Ad Allocation: Interesting Problems Online Stochastic DA Simultaneous online worst-case & stochastic optimization. Tradeoﬀ between delivery penalty and eﬃciency: Covering Constraints? More complex stochastic modeling (drift. seasonality. etc.) .

Outline: Online Allocation Online Stochastic Ad Allocation Online Stochastic Packing Online Generalized Assignment (with free disposal) Experimental Results Online Stochastic Weighted Matching .

Sample a matching Ms from x ∗ . 2.Online Stochastic Weighted Matching “ALG is α-approximation?” if Power of Two Choices Oﬄine: E [ALG(H)] E [OPT(H)] ≥α ∗ 1. 3. . try M . Online: try the edge in Ms ﬁrst. Let M = M1 \Ms where M1 is the maximum weighted matching. Find an optimal fractional solution xe to a discounted 1 matching LP. where xe ≤ 1 − e . and if it doesn’t work.

Thm: Approximation factor is better than 0..66. Sample a matching Ms from x ∗ . try M . . 2.Online Stochastic Weighted Matching “ALG is α-approximation?” if Power of Two Choices Oﬄine: E [ALG(H)] E [OPT(H)] ≥α ∗ 1. Find an optimal fractional solution xe to a discounted 1 matching LP. 2011). Online: try the edge in Ms ﬁrst. where xe ≤ 1 − e . Let M = M1 \Ms where M1 is the maximum weighted matching. (Haeupler. and if it doesn’t work. ZadiMoghaddam. M. 3.

Open Problems

**Online Stochastic Display Ad Allocation
**

Simultaneous online worst-case & stochastic optimization. Tradeoﬀ between delivery penalty and eﬃciency: Covering Constraints? More complex stochastic modeling (drift, seasonality, etc.)

**Online Stochastic Weighted Matching
**

Online Stochastic Matching: Close gap between 0.703 & 0.81. Online Stochastic Weighted Matching: Power of many choices? Better lower bound? Online Weighted Matching (with Free Disposal): Is 1 − 1/e-approximation possible?

Contract-based Ad Delivery: Outline

**Basic Information Ad Serving.
**

Targeting. Online Allocation

Ad Planning: Reservation

Display Ad Delivery: Overview

**Display Ad Delivery Planning:
**

Oﬄine, Online Strategic, Stochastic

Delivery Constraints, Budget

Forecasting

Supply of impressions Demand for ads

Ad Serving: Targeting:

CTR

Allocation:

Online, Stochastic

Ad Planning: Research Issues Which set of contracts should we accept? .

Ad Planning: Research Issues Which set of contracts should we accept? Related Research Issues. Oﬄine Optimization of Contracts. Contracts with Delivery Penalty. . Pricing Uncertain Inventory. Stochastic Supply and Demand Bundling Opportunities Online Mechanisms for Signing Contracts.

di) bi bj Advertisers Impressions Ji di = 4 Jj Eﬃciency (or Revenue) Maximization: Find an assignment with the maximum value. . Needs di impressions (Demand).g.General Ad Planning: Weighted Matching n advertisers. and set Y of impressions (items). bi. Each advertiser i Interested in a set Ji of impressions. (Ji. young women in Seattle). (e. Value vit (or Bid bi ) for each impression t.

Contracts with Delivery Penalty If we don’t meet the demand of an advertiser this month. we should give him/her free impressions in the next month. .

Penalty λbi for not satisfying each unit (Guaranteed Delivery). .Contracts with Delivery Penalty If we don’t meet the demand of an advertiser this month. Each advertiser i Needs di impressions. Bids bi for each impression. we should give him/her free impressions in the next month.

Penalty λbi for not satisfying each unit (Guaranteed Delivery). Bids bi for each impression. (Ji. Needs di impressions.g. di) bi bj Advertisers Impressions Ji di = 4 Jj .Ad Planning With Penalties n advertisers. young women in Seattle). and set Y of impressions (items). bi. Each advertiser i Interested in set Ji of impressions. (e.

and set Y of impressions (items). bi. young women in Seattle). di) bi bj Advertisers Impressions Ji di = 4 Jj Goal: Choose a set T of advertisers to maximize revenue. (e.Ad Planning With Penalties n advertisers. Bids bi for each impression.g. f (T ). Penalty λbi for not satisfying each unit (Guaranteed Delivery). Each advertiser i Interested in set Ji of impressions. Needs di impressions. . (Ji.

Relation to Weighted Matching If we give q items to advertiser i. we get qbi − λbi (di − q) = q(1 + λ)bi − di λbi = q(1 + λ)bi − ci .

Each ad can go to at most one advertiser. P(T ) be the maximum weighted matching in the following bipartite graph with (1 + λ)bi as weights of edges. T Advertisers (1 + λ)bi (1 + λ)bj Impressions P (T ) = Maximum weighted matching for this graph. .Relation to Weighted Matching If we give q items to advertiser i. we get qbi − λbi (di − q) = q(1 + λ)bi − di λbi = q(1 + λ)bi − ci If we commit to a set T of advertisers: f (T ) = P(T ) − C (T ) = i∈T i∈T ci = P(T ) − C (T ). ci where ci = di λbi . Advertiser i ∈ T needs at most di ads.

. maximizing f (T ) is easy as it is a maximum weighted matching.Discussion Summary Given a set T of advertisers.

. maximizing f (T ) is easy as it is a maximum weighted matching.Discussion Summary Given a set T of advertisers.

Discussion Summary Given a set T of advertisers. Can have diﬀerent penalty factors λi for each advertiser i. maximizing f (T ) is easy as it is a maximum weighted matching. Can be used in a negotiation process. . Challenge: Which set of advertisers T should we accept to maximize f (T )? Note: Oﬄine Optimization.

Nazerzadeh. Hardness: No Constant-factor Approximation.Ad Planning with Penalties Feige. At each step. .. Immorlica. Heuristic Greedy Algorithms: Simple Greedy Algorithm: Bicriteria Approximation. M. 2008. add the advertiser with the maximum proﬁt-per-impression ﬁxing the existing assignment.

. At each step. M. Nazerzadeh. add the advertiser with the maximum proﬁt-per-impression ﬁxing the existing assignment. .Ad Planning with Penalties Feige. Heuristic Greedy Algorithms: Simple Greedy Algorithm: Bicriteria Approximation. 2008. Hardness: No Constant-factor Approximation. Theorem: This is a good approximation compared to the optimum with larger penalty factor. Immorlica.

Immorlica. Hardness: No Constant-factor Approximation. At each step. 2008. Nazerzadeh. Heuristic Greedy Algorithms: Simple Greedy Algorithm: Bicriteria Approximation. Theorem: This is a good approximation compared to the optimum with larger penalty factor. add the advertiser with the maximum proﬁt-per-impression ﬁxing the existing assignment. M. ..Ad Planning with Penalties Feige. Greedy-Rate Algorithm: Structural Approximation.

add the advertiser with the maximum marginal-proﬁt per marginal-cost ratio. .Greedy Algorithms Greedy-Rate Algorithm: At each step.

Greedy Algorithms Greedy-Rate Algorithm: At each step. if P(S ∪ {i}) − P(S) − ci > 0. add the advertiser with the maximum marginal-proﬁt per marginal-cost ratio. T : Set of advertisers we committed to (initialized to ∅) At each step. ci . add an advertiser i ∈ X \S to T that maximizes P(S∪{i})−P(S) .

ci Theorem: The greedy-rate algorithm achieves the best structural approximation. add an advertiser i ∈ X \S to T that maximizes P(S∪{i})−P(S) . add the advertiser with the maximum marginal-proﬁt per marginal-cost ratio. . if P(S ∪ {i}) − P(S) − ci > 0.Greedy Algorithms Greedy-Rate Algorithm: At each step. T : Set of advertisers we committed to (initialized to ∅) At each step.

T : Set of advertisers we committed to (initialized to ∅) At each step.Greedy Algorithms Greedy-Rate Algorithm: At each step.. ci Theorem: The greedy-rate algorithm achieves the best structural approximation. 1−α−α ln 1−α 1 α . and f (T ). P(OPT) Greedy-rate algorithm achieves at least factor √ (improving factor 1 + α − 2 α).e. add the advertiser with the maximum marginal-proﬁt per marginal-cost ratio. Proof: Uses submodularity of P(T ). a Signature: α = C (OPT) . add an advertiser i ∈ X \S to T that maximizes P(S∪{i})−P(S) . i. The approximation factor of the algorithm is a function of the structure of the solution. if P(S ∪ {i}) − P(S) − ci > 0.

. • Special case: Demand=1. Fairness. Hartline. Knapsack • Constantin.. Feldman.. Muthkrishnan. Truthful Mechanism: Constant-factor .Ad Planning with Delivery Penalty Banner Ad Delivery Constraints: Guaranteed Delivery. Kleinberg: Online Algorithms with Buyback • Special cases: Single item. Which ads/advertisers should we commit to? FIMN’08 Planning: Online Stochastic Strategic BHK 09 CFMP 09 • Babaiof. Matroid. Pal: Ad slotting with Cancellations.

Hartline. Knapsack • Constantin. Muthkrishnan. Which ads/advertisers should we commit to? FIMN’08 Planning: Online Stochastic Strategic BHK 09 CFMP 09 • Babaiof. Kleinberg: Online Algorithms with Buyback • Special cases: Single item. Truthful Mechanism: Constant-factor Interesting Algorithmic Problem: Online mechanism for general ad planning with delivery penalty: bicriteria approximation? . • Special case: Demand=1. Feldman.. . Pal: Ad slotting with Cancellations. Fairness..Ad Planning with Delivery Penalty Banner Ad Delivery Constraints: Guaranteed Delivery. Matroid.

Tarjan 94 Faster Algorithm for max-ﬂow: running time depends on the size of smaller part.Ad Planning: Oﬄine Optimization of Contracts Fast Algorithms to Verify Feasibility of Contracts: Lopsided Bipartite Graphs. Stein. . Improved Algorithms for Bipartite Network Flow: Ahuja. Orlin.

Orlin. . WHP veriﬁes if there exists a feasible matching. Tarjan 94 Faster Algorithm for max-ﬂow: running time depends on the size of smaller part. Devanur. Chickering. Stein. Improved Algorithms for Bipartite Network Flow: Ahuja.Ad Planning: Oﬄine Optimization of Contracts Fast Algorithms to Verify Feasibility of Contracts: Lopsided Bipartite Graphs. and minimize the future failure probability. Sanghi 2010. Sampling and Concise Allocation. Algorithm: Iteratively assign nodes. Sampling for Max-Cardinality Matching: Charles. Jain.

Sampling and Concise Allocation. Devanur. Algorithm: Iteratively assign nodes. Tarjan 94 Faster Algorithm for max-ﬂow: running time depends on the size of smaller part. Arcuate. Orlin. Stein. Chickering. Malekian. Khuller. Sanghi 2010. Improved Algorithms for Bipartite Network Flow: Ahuja. Jain. and minimize the future failure probability. Online algorithms for accepting contracts: Alaei.Ad Planning: Oﬄine Optimization of Contracts Fast Algorithms to Verify Feasibility of Contracts: Lopsided Bipartite Graphs. Tomlin 2009 Utility model to combine contract-based advertisers & sales-based advertisers. WHP veriﬁes if there exists a feasible matching. Sampling for Max-Cardinality Matching: Charles. Online algorithm for accepting contracts (under assumptions) . Ma.

Other interactions Learning + allocation Learning + auction Auction + contracts .Outline of this talk Ad delivery for contract based settings Planning Ad Serving Ad serving in repeated auction settings General architecture. Allocation for budget constrained advertisers.

Other interactions Learning + allocation Learning + auction Auction + contracts . Allocation for budget constrained advertisers.Outline of this talk Ad delivery for contract based settings Planning Ad Serving Ad serving in repeated auction settings General architecture.

Three main theory/practice problems .

. Find allocation policy to maximize yield = revenue(AdX) + γ · quality(advertisers). where γ ≥ 0 is a tradeoﬀ paramater (or Lagrange multiplier).Combined Allocation with AdX: Objective Short-term: boost revenue from AdX VS Long-term: prioritize quality of guaranteed contracts.

Publisher’s Decisions Assign to an advertiser or discard Impression n-th arrives with quality Qn Submit to AdX with price p Assign to an advertiser or discard ac ce pt re jec t Obtain payment .

Publisher’s Decisions Assign to an advertiser or discard Impression n-th arrives with quality Qn Submit to AdX with price p Assign to an advertiser or discard ac ce pt re jec t Obtain payment Optimal to always test the exchange! .

.Impact of AdX 1000 900 Quality Advertisers 800 700 600 500 400 Pareto efficient frontier ∞ 350 370 390 410 Revenue AdX 430 450 γ = 0: maximum revenue from AdX. γ = ∞: maximum placement quality for contracts.

Online Strategic. Stochastic Delivery Constraints. Budget Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: CTR Allocation: Online. Stochastic .Display Ad Delivery Display Ad Delivery Planning: Oﬄine.

Stochastic Feedback . Budget Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: CTR Allocation: Online. Stochastic Delivery Constraints. Online Strategic.Display Ad Delivery Display Ad Delivery Planning: Oﬄine.

**Online Learning & Allocation
**

Value: Estimated Click-Through-Rate (CTR).

**Online Learning & Allocation
**

Value: Estimated Click-Through-Rate (CTR). Combined online capacity planning & learning?

Budgeted Active Learning

Madani, Lizotte, Greiner 2004, Active Model Selection.

**Online Learning & Allocation
**

Value: Estimated Click-Through-Rate (CTR). Combined online capacity planning & learning?

Budgeted Active Learning

Madani, Lizotte, Greiner 2004, Active Model Selection.

**Bayesian Budgeted Multi-armed Bandits:
**

Guha, Munagala, Multi-armed Bandits with Metric Switching Costs. Goel, Khanna, Null, The Ratio Index for Budgeted Learning, with Applications. Guha, Munagala, Pal, Multi-armed Bandit with Delayed Feedback.

Munagala. Budgeted Unknown-CTR Multi-armed Bandit Pandey. Multi-armed Bandits with Metric Switching Costs. Active Model Selection. Bayesian Budgeted Multi-armed Bandits: Guha. Combined online capacity planning & learning? Budgeted Active Learning Madani. The Ratio Index for Budgeted Learning. Guha. Pal. Greiner 2004.Online Learning & Allocation Value: Estimated Click-Through-Rate (CTR). Null. Lizotte. Handling Advertisement of Unknown Quality. Goel. Olston 2007. Multi-armed Bandit with Delayed Feedback. Munagala. Khanna. with Applications. .

Handling Advertisement of Unknown Quality. Olston 2007. .Online CTR Learning: Mixed Explore/Exploit Pandey.

ni is the number of queries of type i so far. Handling Advertisement of Unknown Quality.Online CTR Learning: Mixed Explore/Exploit Pandey. nia is the number of ˆ times i has been assigned to a. . Algorithm: Revised Greedy Upon arrival of query of type i. assign it to an ad a maximizing ln Pia = (ˆia + 2 niani )bia c where cia is the current estimate of CTR. Olston 2007.

ni is the number of queries of type i so far. assign it to an ad a maximizing ln Pia = (ˆia + 2 niani )bia c where cia is the current estimate of CTR. Algorithm: Revised Greedy Upon arrival of query of type i. opt 2 − O(ln n) where n is the number of . Olston 2007. nia is the number of ˆ times i has been assigned to a. Handling Advertisement of Unknown Quality. Thm[PO07]: ALG ≥ arrivals.Online CTR Learning: Mixed Explore/Exploit Pandey.

.Hybrid ad serving: Contracts + Spot Auctions Given a page view. Auction-based. and two types of advertisers: Contract-based.

Hybrid ad serving: Contracts + Spot Auctions Given a page view. and two types of advertisers: Contract-based. Requirements: For each contract-advertiser. Auction-based. meet its demand. . Implement the scheme using proxy-bidding for contract-advertisers in the spot auction. Decide who wins and how much do they pay.

Hybrid ad serving: Contracts + Spot Auctions Naive solution: If a contract-adv is eligible and has not ﬁnished demand. then let it win the spot. Bid inﬁnity for all auctions. .

Hybrid ad serving: Contracts + Spot Auctions Naive solution: If a contract-adv is eligible and has not ﬁnished demand. then let it win the spot. Bid inﬁnity for all auctions. Optimize for revenue: If the auction pressure (price) is low then let the contract-adv win. . Bid a low bid for all auctions.

then let it win the spot. since low auction-price ⇒ it is a lower value impression. .Hybrid ad serving: Contracts + Spot Auctions Naive solution: If a contract-adv is eligible and has not ﬁnished demand. Bid a low bid for all auctions. Bid inﬁnity for all auctions. Optimize for revenue: If the auction pressure (price) is low then let the contract-adv win. Unfair to contract-adv.

i. Bid a low bid for all auctions. Optimize for revenue: If the auction pressure (price) is low then let the contract-adv win. Bid inﬁnity for all auctions. since low auction-price ⇒ it is a lower value impression. A price-oblivious scheme. an equal slice of impressions from each price-point. bid without seeing the auction bids. Unfair to contract-adv.e..Hybrid ad serving: Contracts + Spot Auctions Naive solution: If a contract-adv is eligible and has not ﬁnished demand. . Ideally: Provide contract-adv with a representative allocation. then let it win the spot. Revenue per auction: average auction-price of impressions given away to contract-advertisers is at most some target t.

Obtaining representative allocations Two main ideas: 1. . Can implement any decreasing function a(p) for fraction of impressions of auction-price p.

Obtaining representative allocations Two main ideas: 1. Can implement any decreasing function a(p) for fraction of impressions of auction-price p. a) s. 2. Solve the system for well chosen distance functions: Minimize dist(U.t.: p a(p)f (p)dp = d pa(p)f (p)dp ≤ td p .

Babaioﬀ.Sharma.Online Learning & Auction Incentives [Devanur.g.Kakade’09. Can these be run in tandem with truthful auctions? (e. UCB). 2nd price for a single slot).Slivkins’09] Multi-Armed Bandit algorithms achieve an “implicit” √ exploration-exploitation tradeoﬀ to get a regret of O( T ) (e.g... .

. A naive explore-exploit method gets O(T 2/3 ) regret: Explore ads for the ﬁrst phase. . 2nd price for a single slot).Online Learning & Auction Incentives [Devanur. Run 2nd price auction for the 2nd phase.Slivkins’09] Multi-Armed Bandit algorithms achieve an “implicit” √ exploration-exploitation tradeoﬀ to get a regret of O( T ) (e. UCB). Can these be run in tandem with truthful auctions? (e.Kakade’09. Babaioﬀ. Fix the CTRs thus learned in the ﬁrst phase.Sharma. giving them out for free.g..g.

. Can you do better that this simpe decoupling? No! Theorem [DK09.Online Learning & Auction Incentives [Devanur. there exist bids. A naive explore-exploit method gets O(T 2/3 ) regret: Explore ads for the ﬁrst phase. s.g. Can these be run in tandem with truthful auctions? (e. ctrs.Slivkins’09] Multi-Armed Bandit algorithms achieve an “implicit” √ exploration-exploitation tradeoﬀ to get a regret of O( T ) (e.Kakade’09.BSS09] For every truthful auction (under certain assumptions).. giving them out for free. Run 2nd price auction for the 2nd phase.. 2nd price for a single slot). Babaioﬀ.t.Sharma.g. Fix the CTRs thus learned in the ﬁrst phase. UCB). regret = Ω(T 2/3 ).

Budget Forecasting Supply of impressions Demand for ads Ad Serving: Targeting: CTR Allocation: Online. Online Strategic. Feature selection and correlation in learning CTR. Stochastic Feedback Feedback Open Problems: Optimal combined online allocation & learning. Stochastic Delivery Constraints.Display Ad Delivery Display Ad Delivery Planning: Oﬄine. Optimal combined stochastic planning and serving? .

Thank You .

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd