First order Gaussian graphs for efficient structure classification

Abstract

First order random graphs as introduced by Wong are a promising tool for structure-based classification. Their complexity, however, hampers their practical application. We describe an extension to first order random graphs which uses continuous Gaussian distributions to model the densities of all random elements in a random graph. These First Order Gaussian Graphs (FOGGs) are shown to have several nice properties which allow for fast and efficient clustering and classification. Specifically, we show how the entropy of a FOGG may be computed directly from the Gaussian parameters of its random elements. This allows for fast and memoryless computation of the objective function used in the clustering procedure used for learning a graphical model of a class. We give a comparative evaluation between FOGGs and several traditional statistical classifiers. On our example problem, selected from the area of document analysis, our first order Gaussian graph classifier significantly outperforms statistical, feature-based classifiers. The FOGG classifier achieves a classification accuracy of approximately 98%, while the best statistical classifiers only manage approximately 91%.