Graph clustering based model building

Abstract

Probabilistic models of high-order statistics, capable of expressing complex variable interactions, have been successfully applied by estimation of distribution algorithms (EDAs) to render hard problems tractable. Unfortunately, the dependence structure induction stage in these methods imposes a high computational cost that often dominates the overall complexity of the whole search process. In this paper, a new unsupervised model induction strategy built upon a maximum flow graph clustering technique is presented. The new approach offers a model evaluation free, fast, scalable, easily parallelizable method, capable of complex dependence structure induction. The method can be used to infer different classes of probabilistic models. © 2010 Springer-Verlag.