Content Based Recommendation and Summarization in the Blogosphere

1
Content Based Recommendation and Summarization in
the Blogosphere
CLAIR

• Ahmed HassanUniversity of Michigan Ann
  ArborDragomir RadevUniversity of Michigan Ann
  ArborJungho Cho University of California Los
  Angeles
• Amruta Joshi University of California Los Angele

2
Outline

• Introduction
• Approach
• Experiments and Results
• Conclusions

1
2
3
4
Introduction Approach Experiments
and Results Conclusions
3
The Blogosphere

• Blogs are now one of the main means for spread of
  ideas and information.
• The size of the Blogosphere has been exhibiting
  an exponential increase.
• How can we decide which blogs are more
  important/influential?

Introduction Approach Experiments
and Results Conclusions
4
The Blogosphere

• Blogs are now one of the main means for spread of
  ideas and information.
• The size of the Blogosphere has been exhibiting
  an exponential increase.
• How can we decide which blogs are more
  important/influential?

Introduction Approach Experiments
and Results Conclusions
5
The Blogosphere

• Blogs are now one of the main means for spread of
  ideas and information.
• The size of the Blogosphere has been exhibiting
  an exponential increase.
• How can we decide which blogs are more
  important/influential?

Introduction Approach Experiments
and Results Conclusions
6
Problem Definition

• Given a set of blogs related to a particular
  topic
• Find a subset of blog feeds to read that have
  interest in that topic
• Similar to the Blog Distillation task in the
  TREC Blog Track.

Introduction Approach Experiments
and Results Conclusions
7
Problem Definition

• Given a set of blogs related to a particular
  topic
• Find a subset of blog feeds to read that have
  interest in that topic
• Similar to the Blog Distillation task in the
  TREC Blog Track.

Introduction Approach Experiments
and Results Conclusions
8
Problem Definition

• Given a set of blogs related to a particular
  topic
• Find a subset of blog feeds to read that have
  interest in that topic.
• Similar to the Blog Distillation task in the
  TREC Blog Track.

Introduction Approach Experiments
and Results Conclusions
9
Ranking Web Pages

• This problem is similar to the problem of ranking
  web pages.
• How is that problem solved?
• Link popularity algorithms
• The most popular link popularity based algorithms
  are
• PageRank
• Hypertext Induced Topic Selection (HITS)

Introduction Approach Experiments
and Results Conclusions
10
Ranking Web Pages

• This problem is similar to the problem of ranking
  web pages.
• How is that problem solved?
• Link popularity based algorithms.
• The most popular link popularity based algorithms
  are
• PageRank
• Hypertext Induced Topic Selection (HITS)

Introduction Approach Experiments
and Results Conclusions
11
Ranking Web Pages

• This problem is similar to the problem of ranking
  web pages.
• How is that problem solved?
• Link popularity algorithms
• The most popular link popularity based algorithms
  are
• PageRank
• Hypertext Induced Topic Selection (HITS)

Introduction Approach Experiments
and Results Conclusions
12
Ranking Web Pages

• Can we use link popularity based algorithms for
  speeches and blogs?
• Yes, but they might not work very well.
• Weakly linked nature of blog pages.
• Bloggers try to exploit the system to boost the
  rank of their blogs.

Introduction Approach Experiments
and Results Conclusions
13
Ranking Web Pages

• Can we use hyperlink popularity based algorithms
  for speeches and blogs?
• Yes, but they might not work very well.
• Weakly linked nature of blog pages.
• Bloggers try to exploit the system to boost the
  rank of their blogs.

Introduction Approach Experiments
and Results Conclusions
14
From Hyperlinks to Similarity

• Use textual similarity to link posts instead of
  hyperlinks.
• Textual similarity between posts suggests some
  kind of interaction
• one of them affecting the other.
• Textual similarity is a way of conferring
  authority.

Introduction Approach Experiments
and Results Conclusions
15
Outline

• Introduction
• Approach
• Experiments and Results
• Conclusions

1
2
3
4
Introduction Approach Experiments
and Results Conclusions
16
Problem Statement

• Given a set of blogs
• Build a graph where
• Nodes represent posts/feeds.
• Edges link posts/feeds with similar text.
• Edge weight reflects how similar the text is.
• Find a subset of nodes that are more important
  than others.

Introduction Approach Experiments
and Results Conclusions
17
Problem Statement

• Given a set of blogs
• Build a graph where
• Nodes represent posts/feeds.
• Edges link posts/feeds with similar text.
• Edge weight reflects how similar the text is.
• Find a subset of nodes that are more important
  than others

Introduction Approach Experiments
and Results Conclusions
18
Text Salience Scores

• We define the importance score of a blog
  recursively in terms of the scores of its
  neighbors as followswhere deg(q) is the
  degree of node q, and adjq is the set of all
  nodes adjacent to q in the network.
• This can be rewritten in matrix notation
  aswhere S (S(p1), S(p2), . . . , S(pN))
  and the matrix B is the row normalized similarity
  matrix of the graph.

Introduction Approach Experiments
and Results Conclusions
19
Text Salience Scores

• We define the importance score of a blog
  recursively in terms of the scores of its
  neighbors as followswhere deg(q) is the
  degree of node q, and adjq is the set of all
  nodes adjacent to q in the network.
• This can be rewritten in matrix notation
  aswhere S (S(p1), S(p2), . . . , S(pN))
  and the matrix B is the row normalized similarity
  matrix of the graph.

Introduction Approach Experiments
and Results Conclusions
20
Measuring Textual Similarity

• Posts similarity is estimated by the cosine
  similarity between a tf-idf vector representation
  of the posts.
• tf The term frequency in some document.
• idf The inverse document frequency.
• tf-idf tf idf
• Cosine similarity the cosine of the angle
  between the tf-idf vectors.
• Other possible similarity measures
• Edit distance.
• Language models.

Introduction Approach Experiments
and Results Conclusions
21
Measuring Textual Similarity

• Posts similarity is estimated by the cosine
  similarity between a tf-idf vector representation
  of the posts.
• tf The term frequency in some document.
• idf The inverse document frequency.
• tf-idf tf idf
• Cosine similarity the cosine of the angle
  between the tf-idf vectors.
• Other possible similarity measures
• Edit distance.
• Language models.

Introduction Approach Experiments
and Results Conclusions
22
Diversity Ranking

• The second important node (B) is quite similar to
  the first.
• Node C is also important but more diverse w.r.t A
  than B.
• How can we select the second node such that it is
  important and as diverse as possible from the
  first selected node?

Introduction Approach Experiments
and Results Conclusions
23
Diversity Ranking

• The second important node (B) is quite similar to
  the first.
• Node C is also important but more diverse w.r.t A
  than B.
• How can we select the second node such that it is
  important and as diverse as possible from the
  first selected node?

Introduction Approach Experiments
and Results Conclusions
24
Diversity Ranking

• Possible solution Remove selected nodes or nodes
  similar to them from the network.
• May disconnect the network
• Better solution Discount nodes that are too
  similar to previously selected nodes.where
  d(p) is a discounting factor that is inversely
  proportional to the similarity between p and the
  selected nodes

Introduction Approach Experiments
and Results Conclusions
25
Diversity Ranking

• Discount nodes that are too similar to previously
  selected nodes.
• where d(p) is a discounting factor that is
  inversely proportional to the similarity between
  p and the selected nodes

Introduction Approach Experiments
and Results Conclusions
26
Adding Priors

• Relation to other nodes might not be the only
  parameter affecting its importance.
• Other attributes that are related to the node
  itself may be involved.
• How can we incorporate this into the
  approachwhere Q(p) is some node quality
  measure, and ß is a trad-off factor.

Introduction Approach Experiments
and Results Conclusions
27
Adding Priors

• Relation to other nodes might not be the only
  parameter affecting its importance.
• Other attributes that are related to the node
  itself may be involved.
• How can we incorporate this into the
  approachwhere Q(p) is some node quality
  measure, and ß is a trad-off factor.

Introduction Approach Experiments
and Results Conclusions
28
Outline

• Introduction
• Approach
• Experiments and Results
• Conclusions

1
2
3
4
Introduction Approach Experiments
and Results Conclusions
29
Experimental Setup

• Data The TREC BLOG06 dataset (Macdonald and
  Ounis 2006)
• 100,649 feeds and 3,215,171 posts covering an 11
  weeks period
• a list of 17,969 spam blogs
• Objective Rank blogs according to their coverage
  of other blogs.

Introduction Approach Experiments
and Results Conclusions
30
Case Study 1 Blogs

• Data The TREC BLOG06 dataset (Macdonald and
  Ounis 2006)
• 100,649 feeds and 3,215,171 posts covering an 11
  weeks period
• a list of 17,969 spam blogs
• Objective Rank blogs according to their coverage
  of other blogs.

Introduction Approach Experiments
and Results Conclusions
31
Evaluation

• Given a rank of blogs, how can we measure how
  good it is?
• Using diffusion models.
• Diffusion models are originally used in social
  networks to model the spread of influence in a
  network.
• The Linear Threshold Model is one of the most
  popular diffusion models.

Introduction Approach Experiments
and Results Conclusions
32
Evaluation

• Given a rank of blogs, how can we measure how
  good it is?
• Using diffusion models.
• Diffusion models are originally used in social
  networks to model the spread of influence in a
  network.
• The Linear Threshold Model is one of the most
  popular diffusion models.

Introduction Approach Experiments
and Results Conclusions
33
Linear Threshold Model

• Each node has a threshold ? in 0,1 selected
  uniformly at random.
• The diffusion process starts with a set of active
  nodes.
• An inactive node v becomes active if

Introduction Approach Experiments
and Results Conclusions
34
Linear Threshold Model
0.6
0.7
0.5
0.2
0.2
0.3
0.1
0.5
0.4
0.4
0.3
0.5
0.2
0.1
0.3
0.5
Example from (Kempe et al, 2003)
Introduction Approach Experiments
and Results Conclusions
35
Linear Threshold Model
0.6
0.7
0.5
0.2
0.2
0.3
0.1
0.5
0.4
0.4
0.3
0.5
0.2
0.1
0.3
0.5
Example from (Kempe et al, 2003)
Introduction Approach Experiments
and Results Conclusions
36
Linear Threshold Model
0.6
0.7
0.5
0.2
0.2
0.3
0.1
0.5
0.4
0.4
0.3
0.5
0.2
0.1
0.3
0.5
Example from (Kempe et al, 2003)
Introduction Approach Experiments
and Results Conclusions
37
Linear Threshold Model
0.6
0.7
0.5
0.2
0.2
0.3
0.1
0.5
0.4
0.4
0.3
0.5
0.2
0.1
0.3
0.5
Example from (Kempe et al, 2003)
Introduction Approach Experiments
and Results Conclusions
38
Linear Threshold Model
0.6
0.7
0.5
0.2
0.2
0.3
0.1
0.5.
0.4
0.4
0.3
0.5
0.2
0.1
0.3
0.5
Example from (Kempe et al, 2003)
Introduction Approach Experiments
and Results Conclusions
39
Evaluation

• The quality of a rank of blogs can be measured
  by
• How many blogs are covered by the first blog in
  the rank?
• How many blogs are covered by the first 2 blogs
  in the rank?
• And so on ..
• The coverage of a set of blogs is simply the
  number of other blogs it activates using the
  linear threshold model
• Initially mark the nodes in the set as active.
• Apply the linear threshold model.
• Count the number of activated nodes.
• Repeat for M times and take the average.
• To compare two blog ranks
• Compare the coverage of the first 1,2,N blogs in
  each rank.

40
Evaluation

• The quality of a rank of blogs can be measured
  by
• How many blogs are covered by the first blog in
  the rank?
• How many blogs are covered by the first 2 blogs
  in the rank?
• And so on ..
• The coverage of a set of blogs is simply the
  number of other blogs it activates using the
  linear threshold model
• Initially mark the nodes in the set as active.
• Apply the linear threshold model.
• Count the number of activated nodes.
• Repeat for M times and take the average.
• To compare two blog ranks
• Compare the coverage of the first 1,2,N blogs in
  each rank.

41
Evaluation

• The quality of a rank of blogs can be measured
  by
• How many blogs are covered by the first blog in
  the rank?
• How many blogs are covered by the first 2 blogs
  in the rank?
• And so on ..
• The coverage of a set of blogs is simply the
  number of other blogs it activates using the
  linear threshold model
• Initially mark the nodes in the set as active.
• Apply the linear threshold model.
• Count the number of activated nodes.
• Repeat for M times and take the average.
• To compare two blog ranks
• Compare the coverage of the first 1,2,N blogs in
  each rank.

42
Baselines

• Random
• Select blogs uniformly at random.
• Heuristic
• Find the most popular blogs by using a fixed
  goodness criteria (e.g. number of posts).
• Greedy
• Start with an empty set M0 .
• Add the node that maximizes the marginal gain in
  subset quality.
• Quality of a subset is measured by the number of
  nodes it covers in the blogs network.
• Very computationally costly O(V4).

Introduction Approach Experiments
and Results Conclusions
43
Baselines

• Random
• Select blogs uniformly at random.
• Heuristic
• Find the most popular blogs by using a fixed
  goodness criteria (e.g. number of posts).
• Greedy
• Start with an empty set M0 .
• Add the node that maximizes the marginal gain in
  subset quality.
• Quality of a subset is measured by the number of
  nodes it covers in the blogs network.
• Very computationally costly O(V4).

Introduction Approach Experiments
and Results Conclusions
44
Baselines

• Random
• Select blogs uniformly at random.
• Heuristic
• Find the most popular blogs by using a fixed
  goodness criteria (e.g. number of posts).
• Greedy
• Start with an empty set M0 .
• Add the node that maximizes the marginal gain in
  subset quality.
• Quality of a subset is measured by the number of
  nodes it covers in the blogs network.
• Very computationally costly O(V4).

Introduction Approach Experiments
and Results Conclusions
45
Results

• Percentage of covered blogs vs percentage of
  selected blogs for BlogRank order, greedy
  algorithm order, number of posts order, and
  random order -Topic Global Warming.

Introduction Approach Experiments
and Results Conclusions
46
Results

• Percentage of covered blogs vs. percentage of
  selected blogs for BlogRank order, greedy
  algorithm order, number of posts order, and
  random order - Topic iPhone.

Introduction Approach Experiments
and Results Conclusions
47
Generalization to Future Data

• Percentage of activated blogs vs. percentage of
  selected blogs for BlogRank order (all data -
  future data),BlogRank-Future order (history data
  - future data), and Greedy-Future order (history
  data - future data) - Topic Global Warming.

Introduction Approach Experiments
and Results Conclusions
48
Generalization to Future Data

• Percentage of activated blogs vs. percentage of
  selected blogs for BlogRank order (all data -
  future data),BlogRank-Future order (history data
  - future data), and Greedy-Future order (history
  data - future data) - Topic iPhone.

Introduction Approach Experiments
and Results Conclusions
49
Priors

• Average normalized number of posts vs. percentage
  of selected blogs for BlogRank order and BlogRank
  with priors order - Topic Global Warming.

Introduction Approach Experiments
and Results Conclusions
50
Priors

• Percentage of covered blogs vs percentage of
  selected blogs for BlogRank order and BlogRank
  with priors order - Topic Global Warming.

Introduction Approach Experiments
and Results Conclusions
51
Similarity vs. Links
52
Outline

• Introduction
• Approach
• Experiments and Results
• Conclusions

1
2
3
4
Introduction Approach Experiments
and Results Conclusions
53
Conclusions

• This work presents a stochastic graph based
  method for recommending blogs to read with
  interest in some topic.
• The proposed method uses content similairty
  instead of hyperlinks to link blogs.
• It addresses issues like diversity ranking and
  biased ranking using some quality priors.
• The method was evaluated using a large real word
  datasets.

Introduction Approach Experiments
and Results Conclusions
54
Acknowledgments
This work is based upon work supported
by the National Science Foundation under
Grant No. 0534323 , Collaborative Research
BlogoCenter Infrastructure for Collecting,
Mining and Accessing Blogs. Any opinions,
findings, and conclusions or recommendations
expressed in this paper are those of the
authors and do not necessarily reflect the views
of the National Science Foundation.
55

• Thanks

Ahmed Hassan hassanam_at_umich.edu