Online Advertisement Campaign Optimization

04/22/23 1

Online AdvertisementOnline AdvertisementCampaign OptimizationCampaign Optimization

Shi ZhongData Mining and Research Group

Yahoo! Inc. Joint work with Weiguo Liu, Shyam Kapur, and Mayank Chaudhary,

published in IEEE/INFORMS SOLI Conference

04/22/23 2

Agenda

Introduction to online advertisingOnline ad campaign optimization problem

Focus: display advertising (i.e., graphical/banner ads)

Approaches and resultsConclusion

04/22/23 3

Yahoo Sponsored Search

Text Ads

04/22/23 4

Google Content Match

Text Ads

04/22/23 5

Display Ads on Yahoo

LREC, 300x250

04/22/23 6

Online Advertising

Text ads Two main categories, a few major players

Sponsored searchE.g., Google search, Yahoo search, Live.com, Ask.com

Content matchE.g., Google adsense, Yahoo YPN

Cost models: CPC Targeting: search query, page content

Display ads Fragmented market Cost models: CPM, CPC, CPA Targeting: content, demo, geo, behavioral, or none

04/22/23 7

Online Ad Campaign Optimization

Netflix, Q4 AdvertisingBudget=$500k,Drive traffic to netflix.com

Google Adwords$250k{dvd rental, online dvd, online movie, …}

Yahoo Display Ads$150k{yahoo top page + LREC, yahoo movie + N, BT=entertainment/movie, …}

DoubleClick$100k{CNN.COM + LREC, IMDB.com + N, …}

Ad Agencies

04/22/23 8

We focus on …

Display advertising campaignsOptimize media buys given a campaign budget and/or campaign objectives

Maximize # conversions/clicks for a given budget Minimize cost for a given number of conversions/clicks

Experiments inside Yahoo Media buys limited to Yahoo products

04/22/23 9

A Campaign Example

A campaign contains multiple lines/productsA line specifies a product from the publisher, a quantity, and a priceA product consists of page location, position, and profile

Page Location Position Targeting Profiles Impressions (thousands)

CPM($)

Run-Of-Personal N Age>=35, Country=US, FreqCap=1 1,476 0.8

Run-Of-Entertainment

SKY Age>=30 3,060 0.89

Run-Of-Movie LREC Age>=30, Country=US, FreqCap=3 3,000 2.72

Run-Of-Network SKY BT=Entertainment, Country=US, FreqCap=1 8,963 0.65

Run-Of-Espanol LREC BT=Entertainment/Movie, Country=US 60 13.51

Run-Of-Maps LREC Age>=30, State=CA, Mon-Fri 7am-10pm 2,000 3.65

04/22/23 10

Quantity and Price Quantity is capped by inventory availabilityPrice is determined by a bidding process

Except for “guaranteed delivery” – for which advertisers have to pay a premium

Higher bid earns higher priority at ad delivery time, thus has a higher probability getting more impressions

04/22/23 11

Optimization Formulation - I

Maximize profit for a given budget

i

iiiqcpmctrrpcq

i

)1000(max}{

ii

ii

iii

iii

iii

CapqBudgetcpmq

Budgetcpmq

cpmqqctrrpc

0

)1(1000

s.t.

qi = # imps for line i, in thousands

ctr = click through ratecpm = cost per thousand imps

rpc = revenue per clickBudget = total budget= max fraction of Budget per line = profit marginCapi = available # imps for line i

04/22/23 12

Optimization Formulation - II

Minimize cost for a desired number of clicks

i

iiqcpmq

i}{min

ii

ii

iii

iii

iii

ci

ii

CapqBudgetcpmq

Budgetcpmq

cpmqqctrrpc

nqctr

0

)1(1000

1000

s.t. nc = desired # clicks

04/22/23 13

Test Results Take a few historical campaigns with Yahoo for some advertiserCompare simulated results from optimization formulation-II with historical campaignsAverage cost saving (for generating same number of clicks) is 26%

History Cost Optimal Cost Saving Campaign 1 $63,503 $38,741 39% Campaign 2 $276,629 $211,472 24% Campaign 3 $376,955 $279,254 26% Total $717,088 $529,468 26%

sounds simple, but …

04/22/23 14

Prepare inputs to optimization engine

Collect/generate product lines Use historical lines of similar advertisers Use data mining techniques to learn “new” lines that are

expected to perform well Use predictive modeling to discover/explore new lines

Estimate CTR for each product Quantity-CPM curve for each product RPC for a given advertiser/business

04/22/23 15

Identify High CTR Segments

Data examplesPage Location Position Age category Geo Location …… Click

Finance N 45-54 CA 0

Autos LREC 30-34 CA 1

……

Finance LREC 35-44 FL 0

Approach:1. Extract frequent segments (with min # impressions) with frequent

itemset mining algorithm

2. Calculate CTR for each segment

3. Check overlap and temporal stability for high CTR segments

04/22/23 16

Identified Segment Examples

Example high-CTR segments• Page:News + Position:LREC + Age:35-54 CTR=0.31%

• Page:Weather + Position:LREC CTR=0.32%

(Baseline average CTR ~ 0.03%)

CTR numbers seen to be stable over timeCPM estimated from most similar historical lines or

Yahoo’s internal pricing system

04/22/23 17

Conclusion

Data mining and optimization work together nicely to enhance campaign effectivenessAn optimized campaign can be very rewardingFurther research

Ad creative optimization Landing page optimization

04/22/23 18

Questions?