Link Prediction Using Supervised Learning

1
Link Prediction Using Supervised Learning

• Presented by
• Mohammad Hasan

Other Members Vineet Chaoji, Saeed Salem, Dr.
Mohammed Zaki Rensselaer Polytechnic Institute
2
Presentation Outline

• Motivation of the work
• Problem Definition
• Dataset Preparation
• Results
• Discussion on the results
• Acknowledgements

3
Motivation

• Predicting future associations in a social
  network
• Finding missing members from hidden groups
• Modeling association patterns over the time
• Finding the force that drive social associations

4
Problem Definition

• Link Prediction in a social group
• u, v are two members
• Apparently, u and v does not associate
• What is the likelihood that they will associate
  in near future?

5
Dataset

• Coauthorship Data
• A true social network that involves scientists
  from a specific subject area.
• The network is formed by
• Institutional affiliations
• Skill and Interest overlaps
• Potential of being Mutually benefitial
• We know, it satisfy
• Small-world phenomenon
• Power-Law distribution
• Availability
• Publicly
• Very small risk of privacy invasion

6
Our Datasets

• BIOBASE Dataset
• Scientific publication data in the area of life
  sciences
• We experimented for the year 1998 - 2002
• Not available publicly
• DBLP Dataset
• Scientific publication data in Computer Science
• We experimented for the year 1985 - 2002
• Available publicly

7
Experimental Steps

• Prepare Labeled Dataset for Supervised Learning
• A binary classification/prediciotn problem Will
  form a link / Will not form a link
• Compute feature values for instance vectors.
• Use classification algorithms to predict links
• Evaluate and Justify
• Features
• Classification Algorithms

8
Preparing Labeled Datasets

• Finding Training Instances
• Define observation/train years
• Randonly select a pairs of scientists, u and v
  only from these years data
• They makes a training instances, if they did not
  associates in these years.
• Assigning Class Labels
• Define test years that chronologically follows
  from the train years
• If u, v has associations in test years, make the
  class label, 1
• Otherwise, the label is -1

9
Finding Features

• We are predicting links in a social graph
• links are graph edges
• Social graph follows small world phenomenon, the
  popular six-degree separation
• So, naturally, a smaller value of minimum
  distance imply a higher change of future
  associations
• So, graph distance is a good feature

10
Finding Features (2)

• Those, who are more connected, are more likely to
  make new associations.
• More connected people has
• Many publications
• Many neighbors
• of publications or of direct neighbors, can be
  features. Obviously, these two features are
  highly correlated

11
Finding Feature (3)

• People associate to be mutually benefitted.
• That requires (identical/completemtary) skill and
  interest.
• Considering identical case, overlap of interest
  is a good feature.
• In the datasets we have, complementary skill did
  not turn out to be a good feature.

12
Feature Categories
13
Feature Evaluation

• Class density distribution approach
• Visualization approach
• ROC approach
• Single feature classification performance
• Accuracy

14
Class density distributions of features
Keyword-match
Shortest-distance
Neighbor-count
Second shortest distance
15
Computing feature ROC by sliding a bar
Precision increases, Recall decreases
distribution
1
-1
Feature Values