Graph partitioning based on link distributions

Existing graph partitioning approaches are mainly based on optimizing edge cuts and do not take the distribution of edge weights (link distribution) into consideration. In this paper, we propose a general model to partition graphs based on link distributions. This model formulates graph partitioning under a certain distribution assumption as approximating the graph affinity matrix under the corresponding distortion measure. Under this model, we derive a novel graph partitioning algorithm to approximate a graph affinity matrix under various Bregman divergences, which correspond to a large exponential family of distributions. We also establish the connections between edge cut objectives and the proposed model to provide a unified view to graph partitioning. Copyright © 2007, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.