Learning Bayesian Network Structures Without Variable Ordering Influence: A Markov Blanket‐Based Approach

ABSTRACT Several methods in the literature address the discovery, from data, of a directed acyclic graph that is the structure of a Bayesian network. This is a challenging task, due to the combinatorial nature of the space of graphs, which grows exponentially with the number of variables. This paper proposes a new method to learn Bayesian network structures, based on Markov blankets. The proposed method, called DMBBN (Dynamic Markov Blanket Bayesian Network), builds a graph from a set of local structures that are induced based on the Markov blanket of each variable of interest. The local structures are combined to generate a single Bayesian network structure, without repetition of nodes and without cycles. The experiments carried out show that DMBBN is promising, especially in large datasets, as it does not depend on a prior ordering of the variables.