Impedance Coupling in Content-targeted Advertising

1
Impedance Coupling in Content-targetedAdvertising

• Berthier Ribeiro-Neto, UFMG, Brazil
• Marco Cristo, UFMG, Brazil
• Paulo Golgher, Akwan, Brazil
• Edleno Moura da SIlva, UFAM, Brazil
• August 2005

2
On Content-targeted Advertising Summary

• Introduction
• Related Work
• Matching Algorithms
• Conclusions

3
On Content-targeted Advertising Introduction
Web Advertising Revenue Growth Comparisons,
1996-2004, USA (Source IAB)
4
On Content-targeted Advertising Introduction
Web Advertising Revenue Growth Comparisons,
1996-2004, USA (Source IAB)
5
On Content-targeted Advertising Introduction

• Search Advertising
• Main Categories
• Keyword-targeted advertising
• Content-targeted advertising

6
On Content-targeted Advertising Introduction
Keyword-targeted Advertising

• Search Advertising
• Main Categories
• Keyword-targeted advertising
• Content-targeted advertising

User query
Paid list
Title
Description
URL (landing page)
7
On Content-targeted Advertising Introduction
Content-targeted Advertising

• Search Advertising
• Main Categories
• Keyword-targeted advertising
• Content-targeted advertising

Triggering page
Paid list
Page content
8
On Content-targeted Advertising Introduction

• Search Advertising
• Main Categories
• Keyword-targeted advertising
• Content-targeted advertising
• Common Elements
• Pay-per-performance
• Relevance
• Advertisers bid on Keywords

9
On Content-targeted Advertising Introduction
Search Advertising Network
Brands Products Services
Performance feedback
Clicks Conversions
Performance-based payment Business
description Ads
Performance-based payment Relevant ads
Relevant info Relevant ads
Content Information need
Audience Information need
10
On Content-targeted Advertising Introduction
Objectives and Contributions

• Objectives
• Improve relevance of matching strategies in
  content-targeted advertising
• Study impact of distinct pieces of evidence and
  classification
• Proposal of new methods
• Study of document classification using link
  information

11
On Content-targeted Advertising Related Work

• Related Work
• Search Advertising
• Relevance Matching
• Ranking
• Fraud Detection
• Feedback Information
• Cross-referencing Information
• Bayesian Networks
• Query Expansion

12
On Content-targeted Advertising Matching
Algorithms Vector Space Model

• Simple strategy consists in matching ad to page
  content

k2
a
w2a
dj
w2j
q
k1
(0,0)
w1j
w1a
13
On Content-targeted Advertising Matching
Algorithms Simple Strategies
Simple Matching Strategies Triggering Page Triggering Page Advertisement Advertisement Advertisement
Simple Matching Strategies original terms require kw title desc kw
AD X X X
KW X X
AD_KW X X X X
ANDKW X X X
AD_ANDKW (AAK) X X X X X
14
On Content-targeted Advertising Matching
Algorithms Experiments

• Ad Collection
• 1,744 advertisers
• 93,972 ads in 2,029 campaigns
• 68,238 keywords
• Test Collection
• 100 pages of a Brazilian newspaper
• Topics include economy, sports, culture, politics
  etc.

15
On Content-targeted AdvertisingMatching
Algorithms Tuning IDF Factors
16
On Content-targeted AdvertisingMatching
Algorithms Simple strategies comparison
17
On Content-targeted Advertising Matching
Algorithms Vocabulary Impedance

• Triggering page belongs to large contextual scope
  than ad
• Pages are about any subject
• Many specific terms
• Ads are concise and more topic restricted
• More general terms
• Triggering page is diffuse
• Multiple topics
• Unimportant topic can offer good opportunity for
  advertising

18
On Content-targeted Advertising Matching
Algorithms Impedance Coupling
Vocabulary Impedance
Triggering Page
Advertisement
keywords
19
On Content-targeted Advertising Matching
Algorithms Impedance Coupling
Vocabulary Impedance
Triggering Page
Advertisement
New Terms
keywords
20
On Content-targeted Advertising Matching
Algorithms Impedance Coupling Model
R
sim(p, dj)
...
...
D0
D1
Dj
Dk
kNN
wij
...
...
T1
T2
Ti
Tm
T3
21
On Content-targeted AdvertisingMatching
Algorithms Ranking comparison
Rank TF-IDF TF-IDF Impedance Coupling Impedance Coupling
Rank term score term score
1 argentina 0.090 wines 0.251
2 obtained 0.047 wine 0.140
3 class 0.036 whites 0.091
4 whites 0.035 red 0.057
5 french 0.031 grape 0.051
6 origin 0.029 bordeaux 0.045
7 france 0.029 acideness 0.038
8 grape 0.017 argentina 0.037
9 sweet 0.016 aroma 0.037
10 country 0.013 blanc 0.036
... ... ... ... ...
35 wines 0.010 - -
... ...
Terms in red are not shared
22
On Content-targeted Advertising Matching
Algorithms Impedance Coupling Strategies
Matching Strategies Triggering Page Triggering Page Triggering Page Advertisement Advertisement Advertisement Advertisement
Matching Strategies original terms new terms require kw title desc kw landing page
AD X X X
KW X X
AD_KW X X X X
ANDKW X X X
AAK X X X X X
AAK_T X X X X X
AAK_EXP X X X X X X
H X X
AAK_H X X X X X X
AAK_EXP_H X X X X X X X
23
On Content-targeted Advertising Matching
Algorithms Experiments

• Ad Collection
• 1,744 advertisers
• 93,972 ads in 2,029 campaigns
• 68,238 keywords
• Test Collection
• 100 pages of a Brazilian newspaper
• Topics include economy, sports, culture, politics
  etc.
• Web Collection
• TodoBR collection
• 5,939,061 pages of br domain

24
On Content-targeted AdvertisingMatching
Algorithms Impact of impedance coupling
25
On Content-targeted AdvertisingMatching
Algorithms Impact of using page pointed by ad
26
On Content-targeted AdvertisingMatching
Algorithms Comparison among all methods
Methods Hits Hits Hits Hits 11-pt Average 11-pt Average
Methods 1 2 3 total score gain ()
H 28 5 6 39 0.026 75.0
AD 41 32 13 86 0.104
AD_KW 51 28 17 96 0.106 1.9
KW 46 34 28 108 0.125 20.2
ANDKW 49 37 35 121 0.153 47.1
AAK 51 48 39 138 0.168 61.5
AAK_H 52 50 46 148 0.191 83.7
AAK_T 65 49 43 157 0.226 117.3
AAK_EXP 70 52 53 175 0.242 132.7
AAK_EXP_H 64 61 51 176 0.253 143.3
27
On Content-targeted Advertising Matching
Algorithms Conclusions

• Automatic methods
• Considerable gains over matching strategies
• High quality content-targeted advertising is
  feasible and practical

28
On Content-targeted Advertising Matching
Algorithms Bayesian Networks

• Represent joint probability distribution
• Nodes ? variables
• Edges ? dependencies
• Chain rule
• Bayes Law

29
On Content-targeted AdvertisingMatching
Algorithms Comparison among all methods
30
On Content-targeted AdvertisingMatching
Algorithms Comparison among all methods