Evaluating the Structure of a Bayesian Network from Data

Dependency graphs are models for representing probabilistic inter-dependencies among related concepts. The nodes of such a graph represent concepts and edges represent probabilistic dependencies among these concepts. Bayesian networks (which consist of graphs and specifications of joint probability distributions) are by far the most investigated class of dependency graphs and their applications span a large range of disciplines. The construction of Bayesian networks based on domain knowledge, evidential data or a combination of both is particularly challenging due to vast number of possible graph structures for a given domain. In this study, we investigate the problem of constructing the structure of a Bayesian network for a domain based on data. We propose two solutions to this problem: an information-theoretical approach and a probabilistic method and we establish the relationship between these two techniques. Finally, we estimate the adequate size of data needed for learning a model such as a Bayesian network structure