Learning Bayesian Networks For Managing Inventory Of Display Advertisements

Max ChickeringMad ScientistLive LabsMicrosoft Corporation

Display Advertisements

AdExpert

Microsoft’s System for DeliveringDisplay AdvertisementsMicrosoft Properties Only7.5 Billion Impressions/Day$1 Billion/Year Revenue

AdExpert: Inventory

Health Pages Top

HealthPagesSide

Inventory consists of impressionsof targetable attributes:

1. Page Groups

Set of pages + position~6000 page groups

2. Geographic Targeting3. Demographic Targeting4. Behavioral Tags

Examples:

• 1M imp of males on the HealthTop page group• 1M imp of sports enthusiast on the AutoSide page group

AdExpert: Selling Inventory

Charge per impressionCost depends on page group and targets

High-touch marketInventory is guaranteed

Guarantees Result InInventory Management Problems

Pricing: How much do we chargeper impression?Remaining Inventory: Can we fill this order?Selection: Do we want to?(something better coming)Delivery: Given that we have overbooked, how do we prioritize orders?

Capacity Prediction

Capacity PredictionCapacity Prediction

Pricing Pricing RemainingRemainingInventoryInventory SelectionSelection DeliveryDelivery

How many Old Males are coming next week?How many Old Males are coming next week?

Capacity Prediction

Example: on a particular page group…Example: on a particular page group…

• Existing order: 1.2M impressions of OldExisting order: 1.2M impressions of Old• New customer wants 1.2M impressions of MalesNew customer wants 1.2M impressions of Males

Can we satisfy new request?Can we satisfy new request?

OldOld MaleMale

1.5M1.5M 1.5M1.5M

YesYes

OldOld MaleMaleNoNo

0.5M0.5M1M1M

0.5M0.5M

Capacity Prediction

OldOld

MaleMale

Autos FanAutos Fan

Sports FanSports Fan

LocationLocation

InvestorInvestor

How Many Old Males Next Week? Age Gender Sports

Old Male No

Young Female No

Old Male Yes p(Age,Gender,Sports)

Capacity PredictionCapacity Prediction==

Volume Prediction Volume Prediction XX

Population PredictionPopulation Predictionp(Age=Old,Gender=Male)p(Age=Old,Gender=Male)

Past Volume

Prediction

Volume PredictionVolume Prediction

Population PredictionPopulation Prediction

Capacity Prediction In Earlier System

Age Gender SportsOld Male Yes

Old Female No

Young Male No

Young Female Yes

Old Male No

Young Female Yes

Young Male Yes

Old Female No

Young Male Yes

RandomRandomSampleSample

NMaleOldNMaleOldp ),(),(

92

Old Male No

Young Female No

Old Male Yes

Not Many TargetsNot Many Targets

New Version Of AdExpert:Increase Targeting

Current System Maxed Out

Earlier system could not handle any more targeting

Competitors adding more targeting

New Demographic Targets

Add Behavioral Targets300 Targets300 Targets

Capacity Prediction From Sample

Age Gender B1 B2 … BN SportsOld Male Yes

Old Female No

Young Male No

Young Female Yes

Old Male No

Young Female Yes

Young Male Yes

Old Female No

Young Male Yes

300 Variables300 Variables

Millions Millions of of

SamplesSamples

x 6000!x 6000!

Compressing Tables With Bayesian Networks

Age Gender SportsOld Male Yes

Old Female No

Young Male YesAge Gender Sports

• One node for each columnOne node for each column• Edges represent probabilistic dependenceEdges represent probabilistic dependence• Each node stores p(node|parents)Each node stores p(node|parents)• Joint probability: product of conditionals:Joint probability: product of conditionals:

p(Age, Sports, Gender)=p(Age) x p(Gender) x p(Sports|Gender)p(Age, Sports, Gender)=p(Age) x p(Gender) x p(Sports|Gender)

• More independence leads to more compressionMore independence leads to more compression

Bayesian networkBayesian network: : Graphical model for representing a joint probability distributionGraphical model for representing a joint probability distribution

p(Age)p(Age) p(Gender|Age)p(Gender|Age) p(Sports | Gender)p(Sports | Gender)

10101414 possible combinations possible combinationsOnly 119,350 parameters Only 119,350 parameters

Bayesian Network For Hotmail PG

Training:Constructing The Model From Data

TrainingTraining

Age Gender B1 B2 … BN SportsOld Male Yes

Old Female No

Young Male No

Young Female Yes

Old Male No

Old Female No

Young Male Yes

Age

Gender

B1

B3

B2

Sports

Efficient “Look up” Algorithms: p(Age=Old, Efficient “Look up” Algorithms: p(Age=Old, Gender=Male)Gender=Male)

Pre-release Validation: Pre-release Validation: Accuracy better than existing systemAccuracy better than existing systemTiming requirements metTiming requirements met

OFFLINEOFFLINE

Bayesian Network: Updating Over Time

Easy to UpdateEasy to UpdateLocal ProbabilitiesLocal Probabilities Ag

e

Gender

B1

B3

B2

Sports

Old Male No

Young Female No

Old Male Yes

Current Status

Capacity Prediction is working wellValuable inventory is still selling outFewer under-delivered targeted orders

Targeting is increasing

Lessons Learned (I)

Include cost of probability “look up” in learning algorithmInclude cost of probability “look up” in learning algorithm

Lessons Learned (II)

Allow “preferred edges” – Some dependences areAllow “preferred edges” – Some dependences areapriori importantapriori important

Age

Gender

B1

B3

B2

Sports

Questions?

© 2006 Microsoft Corporation. All rights reserved.Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation.Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft,and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation.MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.