AdHeat An Influence-based Diffusion Model for propagating hints to match ads - Presentation ppt (1)

1
AdHeat
1 /16

• An Influence-based Diffusion Model
• for
• Propagating Hints to Match Ads

WWW 2010, Hongji Bao, Edward Y. Chang
Google Research, Beijing, China
2
Structure
2 /16

• Background Idea
• AdHeat Influence-based Influence Propagation
• Method
• Hint word generation
• Influential user ranking
• Influence Propagation
• Experiment
• Influence-based Without Propagation vs.
  Content-based
• With Propagation vs. Without Propagation(Influence
  -based)
• Influential users vs. No-influential users
• Conclusion
• AdHeat is Better

3
Background Idea
3 /16

• Traditional Models
• Content-based content of web page -gt keywords
  -gt ads
• User-targeting users information -gt keywords
  -gt ads
• Problems in Social Network
• Data sparsity fake, incomplete, outdated
• Influential users seldom click ads that highly
  relevant to their expertise
• Idea
• Influential users contents and activities are
  attractive to other users
• Users activities represent the interests of the
  user
• Use users activities to produce a list of hint
  words to represent user
• Propagate influential users hint words over the
  social network

4
Method Hint word generation
4 /16

• Hint word generation characterizes each user with
  a list of words.
• LDA(Latent Dirichlet Allocation) was first
  proposed by Blei, Ng and Jordan to model a
  document as a bag of words.
• Finally, every user get a list of (mk) words.
• sorting the words
• use the top j words

output

k words
input
m characteristics
w1
Words generated or viewed by all users In most
recent t days
c1
w1
Processed by PLDA

c2
wk

cm

5
5 /16
Method Influential user ranking

• Constructing Social Network Graph
• A weighted directed graph G(U,E), U users, E
  edges, weight of edge (ui ,uj) useris
  dependence to userj.
• The edge weight from ui to uj is determined by
  two factors.
• The number of interactions from ui to uj
• Each interactions weight
• Normalize all edge-weights to 0,1 by MAX-MIN
  method.

MAX-MIN Method xi (xi-xmin)/(xmax xmin)
6
6 /16
Method Influential user ranking

• Influence Scores Computing
• HITS (Hypertext Induced Topic Selection) have
  been proven to be successful in social network
  analysis tasks.
• Taking the social network graph G as input, the
  HITS algorithm computes two scores for each user
• hub score contribution to propagate information
• authority score contribution to provide
  attractive contents
• HITS algorithm
• h(n1) ?1 (1 - ?)Wcol an
• a(n1) ?1 (1 - ?)WrowT hn
• Influence Scores
• I ah (1 - a)a
• a ? 0,1

• W adjacency matrix of G
• Wcol (row) W with its column(row) normalized
  to sum to one
• ? a reset probability to guarantee the
  convergence
• Iteration ends when
• h(n1) - hn e
• a(n1) - an e
• e is a predefined convergence threshold.

7
7 /16
Method Influence Propagation

• Illustrative Example

8
8 /16
Method Influence Propagation

• Heat Diffusion Model Equation

RH(vi) heat received by node vi DH(vi) heat
diffused by node vi fi(t) the heat of the node
i at time t ?ij the diffusion coefficient
between nodes i and j
9
9 /16
Method Influence Propagation

• Heat Diffusion Model Algorithm

Pseudo Code
Variables
10
10 /16
Experiment

• Dataset
• Google Confucius (QA community)
• Half a million registered users data in one
  month for conducting AdHeat
• 5,000 active users for targeting ads
• Ads data from Google AdSense

11
11 /16
Experiment

• Evaluation Metrics CTR(Click Through Rate)
• A group of experimental users CTR
• CTRds, de Clicksds, de /
  Impressionsds, de
• ds the start day
• de the end day
• Impressionsds, de the number of times that
  ads are shown to the group users during the
  period from ds to de
• Clicksds, de the number of times that users
  in the group click the showing ads
• ds de the CTR of one day
• Use seven-day average CTR to report the results
  in Table 4 and 5 to smooth one days CTR.
• ???????? ??0 6 ??????d?? / 7

12
12 /16
Experiment 1
Influence-based Without Propagation vs.
Content-based

• Figure 2 Shows
• The peak CTR improvement reaches 82 on the fifth
  day.
• Then, the improvement decreases to 51 on the
  final day.
• The average improvement is 66.9.
• Reasons For Improvement Decrease
• A user is unlikely to click on the same ads.
• The users interests might have shifted.

13
13 /16
Experiment 2
With Propagation vs. Without Propagation(Influence
-based)

• Figure 3 Shows
• The peak improvement reaches 102 on day three.
• Decrease after that to 41 on the final day.
• The average improvement is 66.4.
• Propagation helps!
• Reasons For Improvement Decrease
• timeliness of the propagated words.

14
14 /16
Experiment 3
Influential users vs. No-influential users (both
influence-based and without propagation)

• The groups performance in term of content
  contribution was directly proportional to their
  influence. The more influential the group was,
  the more posts they created on Confucius.
• The most influential group has lower CTR than
  that of the least influential group.

15
Experiment Something
15 /16

• In social networks, about matching ads.
• First, no-influential users do not contribute
  enough content so that the traditional
  content-based ad model may be ineffective due to
  information sparsity.
• Second, the propagation of hint words appears to
  be helpful to remedy the information sparsity
  problem.
• Furthermore, correlating ads with information
  that one follows makes AdHeat effective.

16
Conclusion
16 /16

• In summary, the contributions of this paper are
  as follows
• An influence-based advertising model, which
  diffuses hints from influential users to the
  others for matching ads. Experiments Show that
  this influence-based model is both necessary for
  addressing data sparsity problem and improving ad
  relevance.
• AdHeat algorithm, which generates users hints
  and identifies influential users by performing
  social graph analysis, using PLDA and HITS
  algorithms.
• Using a diffusion model to propagate hints to
  improve information density for each user to
  match ads.
• The experimental results show AdHeat to be
  effective, and the results also reveal useful
  insights into understanding influential users
  reactions to ads.

17
Thanks