CS 69191: Masters Seminar
CS 89191: Doctoral Seminar

Spring 2009


Doctoral Student Presentation:
Partitioning for Signed Bipartite Graphs

Victor Lee


Bipartite graphs with signed edges are an interesting special case of graphs that are becoming increasing prominent, because they model social networks that describe a like/dislike or voting behavior. In the case of a voting network, which has applications in both political and consumer behavior, we have voter vertices, object vertices, and edges with value either +1 or -1. Analysts are motivated to detected trends in voter or consumer behavior. While weighted graphs are well-studied, negatively-valued edges pose unique problems because they invalidate many of the common mathematical techniques. We pose the following research question: How can we effectively partition both voter vertices and object vertices according to similar patterns of edge connectivity? We stress that We propose two methods to solve this problem: one based on K-means clustering, and one employing the Expectation-Maximization (EM) probabilistic modeling technique. While these methods are seemingly quite different, we seek to demonstrate similarities in their mathematical foundations. We test the K-means method on a sample of U.S. Senate votes.