About Me

1
About Me

• Swaroop Butala
• MSCS graduating in Dec 09
• Specialization Systems and Databases
• Interests
• Learning new technologies
• Application of technology to financial sectors

2
Coclustering Documents and words using
BipartiteSpectral Graph Partitioning

• Author Inderjit S. Dhillon
• Department of Computer Sciences
• University of Texas, Austin
• Presented by
• Swaroop Butala, Fall 2008

3
Clustering and Current Solutions(1)

• Clustering
• Collection of Objects
• Future Navigation and Searches

4
Clustering and Current Solutions(2)

• Current Solutions
• K-means
• Fuzzy C-means
• Hierarchical clustering
• Document Clustering
• Word Clustering

5
Document Clustering

• Problem
• Vector Space Model
• Extract Unique Content-Bearing Words
• Word by Document matrix
• Existing Solutions
• K-means Algorithm
• Self organized maps
• Computationally Prohibitive

6
Word Clustering

• Basis of documents in which they Co-occur
• Words that typically associate together in
  documents should be associated with similar
  concepts.
• Uses
• Automatic Classification of documents

7
Co-clustering Documents and Words

• Novel Idea
• Duality of word and document clustering
• Use of Bipartite Graphs
• The clustering problem can now be posed as a
  partitioning problem
• Solution
• Spectral Co-Clustering algorithm

8
Bipartite Graph(1)

• No Edges between Words or between Documents

9
Bipartite Graphs(2)
Adjacency Matrix
10
The Partitioning Problem

• Minimum cut vertex partitions in Bipartite Graphs
• Optimal Solution is NPComplete
• Solutions KL and FM algorithms exist
• Spectral Algorithm gives a good global solution
• Better solutions than KL and FM algorithms

11
Graph Partitioning

• To find equally sized vertex subsets such that
  the cut is minimum
• Eigenvectors as optimal partition vectors
• Since the discrete solution is NP complete
• The Bipartitioning Algorithm

12
Conclusions

• A novel idea of Coclustering Words and Documents
  together is proposed
• A real relaxation to optimal solution of
  partitioning is provided
• Algorithm works well on real examples

13
Critique

• Actual motivation for combining document and word
  clustering is not stated
• The solution is not completely optimal since the
  problem of Partitioning is NP complete

14
Questions?