1. Learning Bayesian Networks For Managing Inventory Of Display Advertisements Max Chickering Mad Scientist Live Labs Microsoft Corporation Max Chickering Mad Scientist Live Labs Microsoft Corporation

2. Display Advertisements

3. AdExpert Microsoft’s System for Delivering Display Advertisements Microsoft Properties Only 7.5 Billion Impressions/Day $1 Billion/Year Revenue Microsoft’s System for Delivering Display Advertisements Microsoft Properties Only 7.5 Billion Impressions/Day $1 Billion/Year Revenue

4. AdExpert: Inventory Health Pages Top Health Pages Side Inventory consists of impressions of targetable attributes: 1. 1.Page Groups Set of pages + position ~6000 page groups 2. 2.Geographic Targeting 3. 3.Demographic Targeting 4. 4.Behavioral Tags Examples: 1M imp of males on the HealthTop page group 1M imp of sports enthusiast on the AutoSide page group

5. AdExpert: Selling Inventory Charge per impression Cost depends on page group and targets High-touch market Inventory is guaranteed Charge per impression Cost depends on page group and targets High-touch market Inventory is guaranteed

6. Guarantees Result In Inventory Management Problems Pricing: How much do we charge per 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? Pricing: How much do we charge per 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?

7. Capacity Prediction PricingRemainingInventorySelectionDelivery How many Old Males are coming next week?

8. Capacity Prediction Example: on a particular page group… Existing order: 1.2M impressions of Old Existing order: 1.2M impressions of Old New customer wants 1.2M impressions of Males New customer wants 1.2M impressions of Males Can we satisfy new request? OldMale1.5M1.5MYes OldMale No 0.5M1M0.5M

9. Capacity Prediction Old Male Autos Fan Sports Fan Location Investor

10. How Many Old Males Next Week? Age Gender Sports Old Male No Young Female No Old Male Yes p(Age,Gender,Sports) Capacity Prediction = Volume Prediction X Population Prediction p(Age=Old,Gender=Male) Past Volume Prediction Volume Prediction Population Prediction

11. Capacity Prediction In Earlier System AgeGenderSports OldMaleYes OldFemaleNo YoungMaleNo YoungFemaleYes OldMaleNo YoungFemaleYes YoungMaleYes OldFemaleNo YoungMaleYes RandomSample Old Male No Young Female No Old Male Yes Not Many Targets

12. 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 Targets Current System Maxed Out Earlier system could not handle any more targeting Competitors adding more targeting New Demographic Targets Add Behavioral Targets 300 Targets

13. Capacity Prediction From Sample AgeGenderB1B2…BNSports OldMaleYes OldFemaleNo YoungMaleNo YoungFemaleYes OldMaleNo YoungFemaleYes YoungMaleYes OldFemaleNo YoungMaleYes 300 Variables MillionsofSamples x 6000!

14. Compressing Tables With Bayesian Networks AgeGenderSports OldMaleYes OldFemaleNo YoungMaleYes AgeGenderSports One node for each column One node for each column Edges represent probabilistic dependence Edges 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) More independence leads to more compression More independence leads to more compression Bayesian network: Graphical model for representing a joint probability distribution p(Age) p(Gender|Age) p(Sports | Gender)

15. 10 14 possible combinations Only 119,350 parameters Bayesian Network For Hotmail PG

16. Training: Constructing The Model From Data Training AgeGenderB1B2…BNSports OldMaleYes OldFemaleNo YoungMaleNo YoungFemaleYes OldMaleNo OldFemaleNo YoungMaleYes Age Gender B1 B3 B2 Sports Efficient “Look up” Algorithms: p(Age=Old, Gender=Male) Pre-release Validation: Accuracy better than existing system Timing requirements met OFFLINE

17. Bayesian Network: Updating Over Time Easy to Update Local Probabilities Age Gender B1 B3 B2 Sports Old Male No Young Female No Old Male Yes

18. Current Status Capacity Prediction is working well Valuable inventory is still selling out Fewer under-delivered targeted orders Targeting is increasing Capacity Prediction is working well Valuable inventory is still selling out Fewer under-delivered targeted orders Targeting is increasing

19. Lessons Learned (I) Include cost of probability “look up” in learning algorithm

20. Lessons Learned (II) Allow “preferred edges” – Some dependences are apriori important Age Gender B1 B3 B2 Sports

21. Questions?

22. © 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.