First International Conference on

1
Structural Link Analysis from User Profiles and
Friends NetworksA Feature Construction Approach

• William H. Hsu, Joseph Lancaster, Martin S. R.
  Paradesi, Tim Weninger
• Monday, 26 March 2007
• Laboratory for Knowledge Discovery in Databases
• Kansas State University
• http//www.kddresearch.org/KSU/CIS/ICWSM-20070326.
  ppt

2
Link Analysis in Social NetworksThe K-State
Corpus
3
Outline

• Background, Related Work and Rationale
• Technical Objective Link Mining in Social
  Networks
• Methodology Graph Feature Extraction
• Experimental Results K-State LJMiner Corpus
• Continuing Work Statistical Relational Models

4
Problem StatementLink Mining in Social Networks

• Problem Definition
• Given records of users of weblog or social
  network service
• Discover
• Features of entities users, communities
• Relationships friendship, membership,
  moderatorship
• Explanations and predictions for relationships
• Goals
• Boost precision and recall of link existence
  prediction
• Find relevant features
• Significance Recommendations (Friendship,
  Membership)

5
Related WorkLink Mining

• Getoor and Diehl (2005) - Graphical model
  representations of link structure
• Ketkar et al. (2005) - Data mining techniques vs
  graph-based representation
• Sarkar Moore (2005) - Change in link structure
  across discrete time steps
• Popescul Ungar (2003) - ER model to predict
  links
• Hill (2003), Bhattacharya Getoor (2004)
  Statistical Relational Learning to resolve
  identity uncertainty
• Resig et al. (2004) - Predicting IM online times
  using friends graph degree
• McCallum et al. (2005) - Inferring roles and
  topic categories based on link analysis

6
Rationale

• Limitations of Current State of the Art
• Do not take graph features into account
• Limited ability to select, extract features
• Novel Contribution Link Mining System
• Extracts, computes features of network model
• Towards dependent types for relational link
  mining
• Rationale
• Desired functionality infer new links from old
• Evaluation precision, recall for link existence

7
Outline

• Background, Related Work and Rationale
• Technical Objective Link Mining in Social
  Networks
• Methodology Graph Feature Extraction
• Experimental Results K-State LJMiner Corpus
• Continuing Work Statistical Relational Models

8
K-State Test BedLJMiner Corpus
9
LiveJournal Topology 1Tools and Security Model
10
LiveJournal Topology 2Definitions
11
Outline

• Background, Related Work and Rationale
• Technical Objective Link Mining in Social
  Networks
• Methodology Graph Feature Extraction
• Experimental Results K-State LJMiner Corpus
• Continuing Work Statistical Relational Models

12
Graph Features 1Node, Pair, Link-Dependent
Node-Dependent Features specific to one node
(vertex) within candidate pair
Indegree (v) Target popularity
Indegree (u) Source popularity
Outdegree (u) Source fertility
Outdegree (v) Target fertility
Pair-Dependent Features specific to one
candidate pair of nodes (vertices)
Link-Dependent Features specific to one link
(edge) in directed graph
13
Graph Features 2Node and Pair Features in
LJMiner
14
LJCrawler

• System Design
• Data acquisition client, injector, parser
• Ancillary issues
• Multi-threading
• Distribution
• Storage
• Analytical postprocessing LJClipper, LJStats
• Distinguishing features of LJCrawler
• Results
• 200 users/second maximum, 5 users/second allowed
• Approximately 2 million pages crawled

15
Outline

• Background, Related Work and Rationale
• Technical Objective Link Mining in Social
  Networks
• Methodology Graph Feature Extraction
• Experimental Results K-State LJMiner Corpus
• Continuing Work Statistical Relational Models

16
Network StatisticsGraph Distance
17
Interpretation of Results

• 941-node graph (Hsu et al., 2006) LJCrawler v1
  output
• 1000-4000 node graphs LJCrawler v2 output

18
Outline

• Background, Related Work and Rationale
• Technical Objective Link Mining in Social
  Networks
• Methodology Graph Feature Extraction
• Experimental Results K-State LJMiner Corpus
• Continuing Work Statistical Relational Models

19
Results

• Establishing an Interdisciplinary Research
  Initiative
• K-State / KU / UNL collaboration
• Resources Linguistic Data Consortium
• NIST evaluations
• Involving End Users of Machine Translation
• Document users
• Machine learning, data mining, info extraction
  researchers
• Novel Applications
• Social networks and collaborative recommendation
• Gisting and beyond

20
Continuing Work

• Information Extraction and Intelligent IR
• Learning models for IE ontologies
• Latent semantic analysis
• Machine Learning
• Natural language learning
• Time series learning and understanding
• Relational and first-order models
• Automated Reasoning
• Probabilistic
• Case-based and analogical
• Data Mining and Warehousing
• Grid Computing

21
References

• Knight, K. Whats New in Statistical Machine
  Translation. Invited Talk, International Joint
  Conference on Artificial Intelligence
  (IJCAI-2005), Edinburgh, UK, August, 2005.
• Knight, K. Graehl, J. (2005). An Overview of
  Probabilistic Tree Transducers for Natural
  Language Processing. In Proceedings of CICLing
  2005, p. 1-24.
• Chiang, D. A hierarchical phrase-based model for
  statistical machine translation. In Proceedings
  of the Conference of the Association for
  Computational Linguistics (ACL 2005), p. 263270.
• Koehn, P., Och, F. J., Marcu, D. (2003).
  Statistical Phrase-Based Translation. In
  Proceedings of HLT-NAACL 2003, the Human Language
  Technology Conference of the North American
  Chapter of the Association for Computational
  Linguistics, May 27 - June 1, 2003, Edmonton,
  CANADA.

22
Acknowledgements

• K-State Lab for Knowledge Discovery in Databases
• Vikas Bahirwani
• Tejaswi Pydimarri
• Andrew King
• Social Networks, Graph Theory, Graph Algorithms
• Kirsten Hildrum (IBM T. J. Watson Labs)
• Todd Easton (K-State, Industrial and
  Manufacturing Systems Engineering)
• Machine Learning
• Dan Roth, Cinda Heeren, Jiawei Han (University of
  Illinois at Urbana-Champaign)
• AnHai Doan (University of Wisconsin Madison)

23
Questions and Discussion