Statistical%20Analysis%20of%20the%20Social%20Network%20and%20Discussion%20Threads%20in%20Slashdot

1
Statistical Analysis of the Social Network and
DiscussionThreads in Slashdot

• Vicenç Gómez, Andreas Kaltenbrunner, Vicente
  López
• Defended by Alok Rakkhit

2
Goals

• Understand underlying pattern of communication
• Lead towards efficient techniques to improve
  system performance
• Evaluate Controversy of a thread

3
Why Slashdot?

• Community-based moderation of message boards
• Scoring system
• Thread comments mainly respond to each other
  rather than to article
• Same dataset as previous studies (characterizing
  its size and lifespan)

4
Network Structure

• Filtered out
• Original Poster (if no other involvement)
• Self-replies
• Anonymous posts
• -1 scores
• Topology created in 3 ways
• Undirected Dense
• Undirected Sparse
• Directed

5
Topology Types
6
Network Structure - Expected Features

• One giant cluster containing vast majority of
  users
• Isolated clusters of two to four
• Two orders of magnitude above random
• Small path lengths
• Small maximum distance

7
Degree Analysis

• High variance
• Degree coefficient very small
• Major diff from traditional social networks
• Moderate reciprocity
• Tail of distribution not authors of posts
• Truncated Log-Normal (LN) hypothesis formed much
  better approximation than Power-Law hypothesis

8
Degree Distribution
9
Effects of Score

• Calculated mean score of users with at least 10
  posts
• Found two classes of writers good and average
• Good writers
• Bias in number of comments received
• More replies to their poorly scored posts than
  those of average users

10
Community Structure

• Most pairs have few comments
• Few have very high, up to 108
• Good writers form backbone of network.

11
Agglomerative Clustering
12
Discussion structure

• Radial tree representation used
• High heterogeneity in shape
• Similar mechanism behind their evolution
• Broad first level, wider second level, followed
  by exponential decay
• Decay due to accessibility, new articles
• Branching for level 0 bell shaped, others have
  continuous decrease (LN fit)

13
RADIAL TREES
14
Branching Factors
15
Evaluating Controversy

• Little work done in area
• Other available method involves training a
  classifier for semantic and structural analysis
• Propose using an h-index
• modified from paper output of researchers
• Simple, based of structure alone
• Factors both number of comments and maximum depth
• Tie breaker to thread with fewer comments

16
Impact

• Cited by 11 papers
• Automatic scoring of posts
• Predicting popularity of online content
• What makes conversations interesting
• Comparing volume vs. interaction