Unintrusive Customization Techniques for Web Advertising Marc - - PowerPoint PPT Presentation

▶
unintrusive customization techniques for web advertising
SMART_READER_LITE
LIVE PREVIEW

Unintrusive Customization Techniques for Web Advertising Marc - - PowerPoint PPT Presentation

Dec 08, 2022 113 likes •454 views

Unintrusive Customization Techniques for Web Advertising Marc Langheinrich Atsuyoshi Nakamura Naoki Abe Tomonari Kamba Yoshiyuki Koseki NEC Corporation, C&C Media Research Laboratories, Japan Overview Introduction Ad targeting

slide-1
SLIDE 1

Unintrusive Customization Techniques for Web Advertising

Marc Langheinrich Atsuyoshi Nakamura Naoki Abe Tomonari Kamba Yoshiyuki Koseki

NEC Corporation, C&C Media Research Laboratories, Japan

slide-2
SLIDE 2

NEC Corporation 2

Overview

Introduction

Ad targeting and current methods Targeting with ADWIZ

The ADWIZ System

Architecture and basic interaction The learning process Experimental results

Conclusions

Overview

slide-3
SLIDE 3

NEC Corporation 3

Ad Targeting

Goal

Show advertisement only to desired

target audience

Means

Dynamically select different ad for

each Web site visitor

Targeting Parameters (Examples)

Browser, OS, time of day, country

1.1 Ad Targeting

slide-4
SLIDE 4

NEC Corporation 4

Manual Ad Targeting

Method

Manually define targeting parameters

for each ad

Advantages

Reaches only desired target audience Predictable (How many ads will be

shown?)

Disadvantages

Laborious to setup and maintain

1.1 Ad Targeting

slide-5
SLIDE 5

NEC Corporation 5

Automated Ad Targeting

Method

Neural network learns user interests

Advantages

Fully automated

Disadvantages

User tracking violates privacy Unable to predict number of times an

ad is shown (contract constraints)

1.1 Ad Targeting

slide-6
SLIDE 6

NEC Corporation 6

Targeting with ADWIZ

Automated Targeting

based on search keywords or page

URI

Respects User Privacy

No user tracking necessary

Handles Contract Constraints

Supports minimum number of

displays and other constraints

1.2 Targeting with ADWIZ

slide-7
SLIDE 7

NEC Corporation 7

Control & Data Flow

Content Site User Ad Server

2.1 Control & Data Flow

Request page Request page Return HTML Return HTML Parse Parse Extract parameters Extract parameters Select ad Select ad Return GIF/JPG Return GIF/JPG Display page Display page Request ad image Request ad image

slide-8
SLIDE 8

NEC Corporation 8

Basic Interaction

User searches for "car" User searches for "car"

Keyword-Based Ad Customization Keyword-Based Ad Customization

2.2 Basic Interaction

slide-9
SLIDE 9

NEC Corporation 9

Basic Interaction

System selects a car related advertisement System selects a car related advertisement

Keyword-Based Ad Customization Keyword-Based Ad Customization

2.2 Basic Interaction

slide-10
SLIDE 10

NEC Corporation 10

Basic Interaction

Page-Based Ad Customization Page-Based Ad Customization

System selects a sports related advertisement System selects a sports related advertisement User browses sports section in directory User browses sports section in directory

2.2 Basic Interaction

slide-11
SLIDE 11

NEC Corporation 11

System Components

Content Provider Ad System Advertiser

Content Site User Database Server Learning System Ad Server Administration Server Advertiser

2.3 ADWIZ Architecture

slide-12
SLIDE 12

NEC Corporation 12

Scheduling Ad Displays

  • 1. Select advertisement

graphic to display

  • 1. Select advertisement

graphic to display

  • 2. Set minimum number
  • f necessary displays
  • 2. Set minimum number
  • f necessary displays
  • 3. What is the timeframe

for showing the ad?

  • 3. What is the timeframe

for showing the ad?

  • 4. Any special keyword

you want to reserve?

  • 4. Any special keyword

you want to reserve?

2.4 Administrative Interface

slide-13
SLIDE 13

NEC Corporation 13

Automatically updates every 3, 10 or 30 Minutes Automatically updates every 3, 10 or 30 Minutes

Updating Display Weights

2.4 Administrative Interface

slide-14
SLIDE 14

NEC Corporation 14

List of ads and their probabilities of being displayed for a certain keyword List of ads and their probabilities of being displayed for a certain keyword

Inspecting the Weights

2.4 Administrative Interface

slide-15
SLIDE 15

NEC Corporation 15

Inspecting the Weights

List of keyword weights per advertisement List of keyword weights per advertisement

2.4 Administrative Interface

slide-16
SLIDE 16

NEC Corporation 16

Inspecting the Weights

List of page weights per advertisement List of page weights per advertisement

2.4 Administrative Interface

slide-17
SLIDE 17

NEC Corporation 17

Inspecting the Weights

List of advertisement weights per page List of advertisement weights per page

2.4 Administrative Interface

slide-18
SLIDE 18

NEC Corporation 18

Keyword based Learning

Advertisements Aj Required displays hj Toyota Camry Toyota Camry Cyberwing Golf Cyberwing Golf 110 000 110 000 50 000 50 000 Keywords Wi Usage rate ki car car golf golf 17 462 17 462 34 921 34 921 Click-through rate cij car golf Toyota Camry Cyberwing Golf 7% 7% 8% 8% 1% 1% 11% 11%

Inputs

∑ ∑

= = m i n j ij i ij d

k c

1 1

Maximize expected total click-through rate

  • 1. Show all required displays
  • 2. Weights sum up to 100%
  • 3. No negative weights allowed

∑ =

=

n i j ij i

h d k

1

∑ =

=

m j ij

d

1

1 ≥

ij

d

2.5 The Learning Process

slide-19
SLIDE 19

NEC Corporation 19

Keyword based Learning

Advertisements Aj Required displays hj Toyota Camry Toyota Camry Cyberwing Golf Cyberwing Golf 110 000 110 000 50 000 50 000 Keywords Wi Usage rate ki car car golf golf 17 462 17 462 34 921 34 921 Click-through rate cij car golf Toyota Camry Cyberwing Golf 7% 7% 8% 8% 1% 1% 11% 11%

Inputs

∑ ∑

= = m i n j ij i ij d

k c

1 1

Maximize expected total click-through rate

  • 1. Show all required displays
  • 2. Weights sum up to 100%
  • 3. No negative weights allowed

∑ =

=

n i j ij i

h d k

1

∑ =

=

m j ij

d

1

1 ≥

ij

d

Output

car golf Display rate dij Toyota Camry Cyberwing Golf

Total:

74% 74% 26% 26%

100%

91% 91% 9% 9%

100%

2.5 The Learning Process

slide-20
SLIDE 20

NEC Corporation 20

Ad Server Ad Server Ad Server

Ad Selection Process

Administration Server Administration Server Learning System Learning System HTTP HTTP

"car" P(Ai|"car") Ai

Database Server Database Server Extract Keyword Lookup Weights Select Ad Return GIF/JPG

Click-Through & Usage rate Weights Required Displays

2.5 The Learning Process

slide-21
SLIDE 21

NEC Corporation 21

Experimental Setup

Keyword based

32 Ads 128 Keywords

Setup

Simulated keyword

search

Artificial User Interest

Models

Repeated 1 million

times

Averaged over 5 runs

Always select the advertisement which had the highest click-through rate for given keyword in the past Always select the advertisement which had the highest click-through rate for given keyword in the past

Methods compared

Random Selection Constraint-based

Learning

Max-Click Method

2.6 Experiments

slide-22
SLIDE 22

NEC Corporation 22

  • Random Selection

Experimental Results

Advertisement ID Number of times

2.6 Experiments

slide-23
SLIDE 23

NEC Corporation 23

  • Random Selection

Experimental Results

Advertisement ID Number of times

  • Max-Click Method
  • More total clicks
  • Fails to show more than

half of the ads

2.6 Experiments

slide-24
SLIDE 24

NEC Corporation 24

  • Random Selection

Experimental Results

Number of times Advertisement ID

  • Constraint-based Learning
  • Increases click-through for

all ads

  • Shows minimum number of

required displays

Advertisement ID Number of times

  • Max-Click Method
  • More total clicks
  • Fails to show more than

half of the ads

2.6 Experiments

slide-25
SLIDE 25

NEC Corporation 25

Max-Click and Random Method identical Max-Click Method better than Random Method Random Method better than Max-Click Method

  • Random Selection

Experimental Results

Number of times Advertisement ID

  • Constraint-based Learning
  • Increases click-through for

all ads

  • Shows minimum number of

required displays

Advertisement ID Number of times

  • Max-Click Method
  • More total clicks
  • Fails to show more than

half of the ads

2.6 Experiments

slide-26
SLIDE 26

NEC Corporation 26

Experimental Results II

Advertisement ID Number of times

Random Method

2.6 Experiments

slide-27
SLIDE 27

NEC Corporation 27

Experimental Results II

Advertisement ID Number of times

Max-Click Method Random Method

2.6 Experiments

slide-28
SLIDE 28

NEC Corporation 28

Experimental Results II

Advertisement ID Number of times

Constraint-Based Learning Max-Click Method Random Method

2.6 Experiments

slide-29
SLIDE 29

NEC Corporation 29

Experimental Results II

Advertisement ID Number of times

Constraint-Based Learning Max-Click Method Random Method

2.6 Experiments

  • More than 15%

improvement over Max-Click Method

slide-30
SLIDE 30

NEC Corporation 30

Conclusions

Current Ad Targeting Solutions

Manual:

Laborious

Automated:

Threatens privacy Difficult to incorporate contract constraints

ADWIZ

Offers Automated Targeting Respects User Privacy Handles Contract Constraints

3.1 Conclusions

slide-31
SLIDE 31

NEC Corporation 31

Future Work

Scaling Up

Thousands of keywords, pages, ads Clustering techniques

Faster Learning for New Ads

How to reuse previously learned

parameters for new advertisements

Real-World Deployment

"Real" experiments

3.2 Future Work

slide-32
SLIDE 32

NEC Corporation 32

Related Work

Web Advertisement

effectiveness [Risden98] alternative forms [Kohda96, Briggs97] customization [Baudisch97]

Privacy

user surveys [Rogers98, Cranor99] cookies & profiling FTC reports, EU Directive

3.3 Related Work

slide-33
SLIDE 33

NEC Corporation 33

ADWIZ Homepage

http://www.ccrl.com/adwiz/

For More Information