Use of Content Tags in Managing Advertisements for Online Videos

1
Use of Content Tags in Managing Advertisements
for Online Videos

• Chia-Hsin Huang, H.T. Kung, Chia-Yung Su
• Harvard School of Engineering and Applied
  Sciences

IEEE CEC/EEE 08 Washington DC 7/24/2008
2
Outline

• Background
• Outline of Approach
• Experimental Results
• Conclusion

3
Motivation Rapid Growth ofOnline Shared Videos

• E.g., over 150,000 new videos are uploaded to
  YouTube everyday

Data Source http//ksudigg.wetpaint.com/page/You
TubeStatistics?tanon
4
Motivation Projected Growth of Online
Advertising in United States
Data source http//www.marketingcharts.com/
5
Goal of This Work Selecting Relevant Ads for
Online Shared Videos

• We want to select relevant ads automatically for
  a given video
• Ideally, use video data analysis to inform ad
  selection
• However, unlike text-based content, extracting
  such information from video data is difficult
• We exploit video tags for selecting relevant ads
• Tags are in general author-generated metadata of
  video content
• Most online shared videos have tags

6
Keyword-based Approach for Selecting Relevant
Advertisements

• It is the most popular approach for Internet
  advertising (e.g., Google AdWords)
• We could use this approach to select ads for
  videos by matching video tags to ad keywords.
  That is

Algorithm
Algorithm
Most Relevant Ad
Video
Find ads in DB where any keyword matches any
video tag
Select the most relevant ad
Candi- date Ads
Tag1 Tag2
Database of Ads
7
Vocabulary Impedance Problem of Keyword-based
Approach

• Even if an ad is related to a video, it may not
  be considered as a candidate unless it has
  keyword matches

Resulting Video Tags
User Search Terms
YouTube Search
PGA Tiger
PGA Tour
Ad Keywords
Golf Clubs
(semantic matching can alleviate the problem
somewhat)
8
Our New Approach Use VideoTags For Selecting
Relevant Ads

• Keyword-based approach suffers from the
  vocabulary impedance problem because advertisers
  usually can only assign ad keywords matching a
  subset of tags of all relevant videos
• We propose an approach which consider all tags in
  all known relevant videos for an ad

8
9
However, Not All Tags Are Important to Advertisers

• Example Different relevance of tags in YouTube
  video Miley Cyrus - 7 Things - Fan Video with
  respect to the ad Miley Cyruss latest album
• This means that we must consider the different
  levels of relevance when we use tags

10
Contributions of This Paper

• We propose an ad selection approach for online
  shared videos based on video tags
• We propose a method of learning the relevance of
  tags to each individual ad
• Our experimental results show that our approach
  can select relevant ads

11
Outline

• Background
• Outline of Approach
• Experimental Results
• Conclusion

12
Problem Definition

• Given
• A large set of videos
• A set of advertisements
• A score sheet indicating how good a subset of
  videos (i.e., training videos) match each ad.
  This is the ground truth
• Objective
• For an arbitrary video, find the most relevant ad
  based on video tags

13
Two-Phase ApproachTo Relevant Ad Selection

• Phase 1 Training
• We train a model describing the relevance of each
  tag to each ad
• Phase 2 Ad Selection
• We use the model to select the most relevant ad
  for a given video

14
Phase 1 Constraining the Model

• We assume a linear relationship between tag
  relevance and ground truth score

Derived Relevance Constraints
Training Videos
Ground Truth Scores
Ad
Tags
T1,T2
9
R1 R2 9
R1 R2 7
7
T1,T2
R1 5
5
T1
15
Phase 1 Solutions To Constraint Equations
Estimate Tag Relevance
Tag Relevance Estimates
Constraint Equations
Ad
R1 R2 9
R1 5
Solve
R1 R2 7
R2 3
(linear regression)
R1 5
16
Phase 2 Ad Selection
Given Video
Estimated Tag Relevance
Select Highest Score Estimates
Ad Set
R1,1 5 R2,1 3 R3,1 1
R1,1 R2,1 8
R1,2 4 R2,2 3 R3.2 2.5
R1,2 R2,2 7
Tags T1,T2
R1,3 2 R2,3 1.5 R3,3 4
R1,3 R2,3 3.5
17
Similarity to Perceptron

• Our model can be viewed as a perceptron-like
  algorithm (n input nodes, m summation nodes)
• Input bit vector of length n, each bit
  represents presence of a tag
• Output m numbers describing the estimated score
• We can thus borrow theory from work on perceptron
  (e.g., Freund 1999)

Estimated Scores
Ads
A1
A2
A3

Tag Relevance
R1,1
R5,3


T1
T2
T3
T4
T5

Tags
Bit vector of tags in a video
18
Outline

• Background
• Outline of Approach
• Experimental Results
• Conclusion

19
Video and Ad Dataset in Experiments

• Video collection process
• Search for Wii on YouTube
• Filter out low view count videos
• Remove stopwords (e.g., the and of) and
  redundancies in video tags
• Remove videos containing only one tag
• 300 YouTube videos collected
• Split 150 for training, 150 for testing
• 64 ads from TV stations, e.g.

20
Score Sheet in Experiments

• We manually generate the score sheet, which
  contains 300 x 64 ad-video relevance scores
• Scores range from 1 to 9
• The scores are used as ground truth

Videos

Ads
6
9
3
4
2
1
7
5
3
3
8
6
21
Evaluation Metric Rank of Selected Ad

• Step 1. For a given video, sort the ad set
  according to the ground truth score, in
  descending order
• Step 2. Use the trained model to select an ad for
  this video
• Step 3. The location of the selected ad in the
  sorted ad set is called the rank of selected ad

Sorted Ad Set for Video X
Rank of Selected Ad (Ad5) is 2
Selected Ad Ad5
Video X
22
Experimental Result 1 Two Populations of Videos
Revealed
Population 1 Highly relevant ads selected
Population 2 Less relevant ads selected due to
rare/few tags
23
Experimental Result 2 Small Training Set Still
Yields Good Results
Doubling training set size contributes only 20
performance gain
Smallest training set can still select highly
relevant ad with a probability of 60
24
Outline

• Background
• Outline of Approach
• Experimental Results
• Conclusion

25
Conclusion Future Work

• We proposed a method for selecting relevant ads
  to a given online shared video using tags
• Our simple linear model can still select relevant
  ads
• 80 of suggested ads are ranked among top 5 when
  using a large training set
• 60 of suggested ads are ranked among top 5 when
  using a small training set
• Future work
• Theoretical analysis on convergence of the method
• Analyze the characteristics of tags which will
  make the method work better

26

• Thank You for Your Attention
• Comments and Questions?

27
Detail of AdWords
Rank Cost-Per-Click Quality Score
Quality Score f(CTR, Keyword and Ad relevance,
Landing Page quality)
28
Properties of tagging

• Most users use few (
• Golder 2005 Marlow 2006
• Power law distribution of tag frequency
• Kipp 2006 Halpin 2007
• People tag the same content using limited number
  of tag combinations
• Golder 2005 Kipp 2006 Halpin 2007
• Conclusion
• We can extract enough tagging information by
  investigating limited number of videos

29
Intuition of Using Tags

• We should consider intrinsic values of tags
• We can find tagging consensus by investigating a
  limited set of tagged content