1 Internet Advertising Ramana Yerneni, Yahoo! Labs [email protected] August 17, 2010.

1

Internet Advertising

Ramana Yerneni, Yahoo! Labs

[email protected]

August 17, 2010

2

Advertisers and Publishers

• Goal of advertisers: reach the target users in a relevant context

• Goal of publishers: derive revenue by presenting ads, along with content, to target users

3

Advertising Media

• Variety of media• Print: Ads in newspapers, magazines, …

• TV: Show ads on TV programs

• Internet: Display ads on Web pages

• Offline vs. Online advertising• Print and TV are considered offline

• Internet advertising is considered online

4

Search vs. Display Advertising

• Keyword search• Strong signal of user intent

• Show ads along with search results

• Advertisers set up campaigns by indicating target keywords

• Ad opportunities are characterized by the keywords specified by search users

5

Search vs. Display Advertising

• Display ads• User profiles and content context are crucial

• Display ads embedded in content pages

• Advertisers specify target user profiles and the desired content context

• When a user visits a Web page, the ad opportunity is characterized by the profile of the user and the content context of the Web page

6

Ad Opportunities

• Specified by attribute values• User information (e.g., Gender, Age)

• Content information (e.g., Web page)

• Context information (e.g., Zip code)

• Example:• Content = finance, Gender = male, Age = 21-

30, Zip = 95051, Platform = IPhone, Interest = sports, …

7

Ad Campaigns

• Specify a constraint that each qualifying ad opportunity must satisfy

• (Attr1 = value1 and Attr2 = value2) or (Attr3 = value3 and Attr4 != value4) …

• Examples:• Gender = male and Zip = 95051

• Age = 31-40 and Zip != 95051

8

Campaign Specification

• Duration is an important constraint• Start and end timestamps are specified

• Volume goals• Campaigns typically specify number of ad

opportunities they seek

9

Setting up a Campaign

• Determine inventory for specified ad campaign

• Forecast ad opportunities that match the campaign

• Factor in contention from other existing campaigns

• Compute the price of the targeted ad opportunities

• Book campaign• If sufficient inventory available, set up the ad

campaign in the system

10

Serving Ads to Campaigns

• An ad opportunity arises when a user visits a Web page

• Identify the set of matching campaigns• Attribute values in ad opportunity satisfy the

constraints specified by these campaigns

• Select a campaign, and an ad to serve• Based on campaign goals for volume of

impressions, budget considerations, etc.

• Deliver the ad and log the event

11

Display-Advertising System

CampaignSetup

Campaign List

Ad Logs

AdServing

Ad Opportunity

Campaign Ad

Campaign Query/Booking

Query/Booking Results

12

Outline

• Overview

• Inventory Forecasting

• Campaign Matching

• Conclusion

13

Inventory Forecasting

• Need to forecast future ad opportunities that match targeted ad campaigns

• First step in the campaign setup process

• Once the set of matching opportunities are forecast, the available inventory, its allocation and pricing aspects are determined

14

Forecast Accuracy

• Over-forecasting causes failure to execute on advertiser campaigns

• Penalties for not fulfilling the campaign goals

• Under-forecasting leads to loss of revenue for publisher

• Not being able to monetize all ad opportunities

15

Performance Requirements

• Real-time considerations• Latency requirements of the order of 100 ms

• Inventory space is enormous• Trillions of ad opportunities in play

16

Projecting from History

• Time-series analysis• Start with logs of ad opportunities

• Generate historical time-series

• Project onto the future

• Trend factors• Weekly patterns: day-to-day variance within a

week

• Seasonal variations: e.g., Christmas shopping

17

Computational Challenges

• Set of inventory queries not known a priori

• Enormous number of possible queries• Precomputation time is too large

• Storage of time series information prohibitively expensive

18

Solution Approach

• Narrow the space of profiles for time-series computation to a subset of attributes

• Derive forecast for advertiser queries on the fly

• Example• Given query: Content = finance and Gender = male and

Duration = [12/01/10 – 12/31/10]• Forecast inventory for core profile: Content = finance

and Duration = [12/01/10 – 12/31/10]• Derive forecast for given query by using a scaling factor

for Gender = male (perhaps 0.5)

19

Forecast Derivation

• Scaling-factor computation for attribute conditions beyond core profiles

• Scaling factors sensitive to core profiles• Consider scaling factor for Gender = male• 60% of users on sports sites may be males• 40% of users on shopping sites may be males • For a given query “Content = finance and Gender

= male and Duration = [12/01/10 – 12/31/10]”, if we have time series for “Content = finance and Duration = [12/01/10 – 12/31/10]”, what scaling factor should be employed?

20

Scaling Factors based on Sampling

• Correlation across multiple conditions• Deriving forecasts for multiple conditions (e.g.,

Gender = male and Age = 31-40) is tricky

• Using simplistic independence assumptions leads to significant errors in computing scaling factors

• Sampling ad opportunities to represent correlation

• Scaling-factor computation based on sample matches for full query and for core-profile query can be much more accurate

21

Forecast-Computation Flow

• Offline computation• Precompute time series, based on historical logs,

for core profiles• Generate sample ad opportunities to enable

forecast derivation

• Online query processing• Generate forecast for the query’s core profile• Using sample ad opportunities for full-query and

core-profile, determine scaling factor• Determine forecast for the query based on the

core-profile forecast and the scaling factor

22

Forecasting System Diagram

GenerateForecastModel

(offline)

Ad Logs

Core-Profile Time Series

Opportunity Samples

ProvideInventoryForecast(online)

Inventory Query

Forecasted Inventory

23

Performance Challenges

• Offline computation• Billions of ad opportunities per day: time-series

computation is complex and arduous

• Solution approach: incremental computation of time series

• Online query processing• Need large number of sample opportunities to

cover a large space of queries

• Solution approach: memory-resident bit-vector indexing, partitioned across multiple servers

24

Accuracy Challenges

• Opportunity sampling• Highly-targeted queries may have few or no

samples

• Time-series computation• Low-volume core profiles can have significant

variance in future projection

• Forecast derivation• Large “distance” between inventory queries and

corresponding core profiles can lead to significant forecast errors

25

Outline

• Overview

• Inventory Forecasting

• Campaign Matching

• Conclusion

26

Ad Opportunities

• Attribute values for user, content and context information

• Also included are attributes like timestamp

• Example • Timestamp = 12/12/10 10:30:15, Content =

finance, Gender = male, Age = 21-30, Zip = 95051, Platform = IPhone, Interest = sports, …

27

Campaign-Opportunity Matching

• Matching campaigns have constraints on attribute values

• Campaigns typically specify a few constraints

• Example• Duration = [12/01/10 – 12/31/10]; Content =

finance and Gender = male and Age = 21-30

28

Campaign-Matching Challenge

• Matching a campaign to an opportunity is a complex operation

• Campaign constraints can be complex Boolean expressions

• Ad opportunities can have large number of attributes

• Low-latency requirement• Given an opportunity, need to identify all matching

campaigns fast (within a few milliseconds)

• Large number (millions) of campaigns in play

29

Naive Solution

• Scan all campaigns• For each campaign, examine if it matches the ad

opportunity

• Problem• Takes too long: large number of attributes to

match; large number of campaigns to examine

• Typically, cannot scale beyond a small number of (say 100) campaigns, with the low-latency requirement

30

Semi-naïve Solution

• Scan a relevant subset of campaigns• Set up buckets for each attribute-value condition (feature)

that appears in some campaign

• Populate the buckets with campaigns that specify the attribute-value condition

• Scan the campaigns in the buckets that match the attribute values in the ad opportunity

• Problem• each opportunity can specify a large number of (say 100)

attribute values; each feature bucket can have a large number of campaigns (e.g., Gender = male)

31

Campaign-Indexing Insights

• Most campaigns have few constraints

• Most campaigns specify constraints in simple conjunctions

• Indexing campaigns that are conjunctions can be done efficiently

• Can easily extend the solution to cover DNF-structured campaigns

32

Inverted Indexing

• Set up buckets for features and identify the buckets that match the ad opportunity

• March down the lists of campaigns in these buckets in an organized manner

• Order each list by campaign id

• Skip through each list to find campaigns that appear in the target number of lists

33

Feature Buckets

Campaign C1 : Content = finance and Age = 31-40

Campaign C2 : Content = sports and Age = 31-40

Campaign C3 : Content = finance and Age = 31-40 and Zip = 95051

Campaign C4 : Content = finance and Age = 21-30

Content = finance

Age = 31-40

Zip = 95051

Age = 21-30

Content = sports

C1

C1 C2

C2

C4

C4

C3

C3

C3

34

Opportunity Query

Campaign C1 : Content = finance and Age = 31-40

Campaign C2 : Content = sports and Age = 31-40

Campaign C3 : Content = finance and Age = 31-40 and Zip = 95051

Campaign C4 : Content = finance and Age = 21-30

Content = finance

Age = 31-40

Zip = 95051

Age = 21-30

Content = sports

C1

C1 C2

C2

C4

C4

C3

C3

C3

Opportunity : Content = finance, Gender = female, Age = 31-40, Zip = 95051, Platform = IPhone

Match: {C1, C3}

Union: {C1, C2, C3, C4} has false positives

Intersection: {C3} has false negatives

35

Postprocessing Computation

• Not all constraints of campaigns are easy to index

• E.g., the Duration constraint

• First, compute the set of campaigns satisfying the indexed constraints

• Then, process other constraints in a postprocessing phase

36

Campaign-Matching System

GenerateCampaign

Index(offline)

Campaign List

Meta data

MatchCampaigns

(online)

Ad Opportunity

Matching Campaigns

Feature buckets

37

Campaign-Matching Challenges

• Scale• Number of attributes in impressions • Number of campaigns

• Constraint complexity• Number of attributes in campaign constraints• Complexity of the boolean expressions in

constraints

• Complexity of the conditions• Range and other operators• Confidence intervals

38

Outline

• Overview

• Inventory Forecasting

• Campaign Matching

• Conclusion

39

Summary

• Focus area: Online/Internet Display Advertising

• Modeling ad campaigns and ad opportunities

• Forecasting inventory for campaigns

• Allocation and pricing aspects

• Matching opportunities and campaigns

40

References

• Forecasting High-Dimensional Data • Proceedings of SIGMOD 2010

• Indexing Boolean Expressions• Proceedings of VLDB 2009

• Adaptive Bidding for Display Advertising• Proceedings of WWW 2009

41

Thank You

Q & A