Mieczysław Alojzy Kłopotek

On the Distance Hypothesis
in Tree-like Bayesian Networks

952

Abstract

Bayesian networks have many practical applications due to their capability to represent joint probability distribution in many variables in a compact way. Though there exist many algorithms for learning Bayesian networks from data, they are not satisfactory because the learned networks usually are not suitable for reasoning. So far only for tree-like and poly-tree Bayesian networks and also for so-called Structured Bayesian networks a satisfactory reasoning algorithms applicable directly for Bayesian networks have been invented.

This radically increases the need for efficient learning algorithms for these classes of Bayesian networks. In fact, algorithms learning tree-like Bayesian networks have been created allowing for learning in case of large numbers of variables. The fastest algorithm, however, relies on the assumption of special node similarity measure properties.

This paper defines and explores a definition of such a similarity measure.

It is also demonstrated that this measure facilitates development of algorithms for learning Structured Bayesian Networks from data.