An information-geometric approach to learning Bayesian network topologies from data

Abstract

This work provides a general overview of structure learning of Bayesian networks (BNs), and goes on to explore the feasibility of applying an information-geometric approach to the task of learning the topology of a BN from data. An information-geometric scoring function based on the Minimum Description Length Principle is described. The info-geometric score takes into account the effects of complexity due to both the number of parameters in the BN, and the geometry of the statistical manifold on which the parametric family of probability distributions of the BN is mapped. The paper provides an introduction to information geometry, and lays out a theoretical framework supported by empirical evidence that shows that this info-geometric scoring function is at least as efficient as applying BIC (Bayesian information criterion); and that, for certain BN topologies, it can drastically increase the accuracy in the selection of the best possible BN. © 2008 Springer-Verlag Berlin Heidelberg.