Learning Robust Bayesian Network Classifiers in the Space of Markov Equivalent Classes

Restricted Bayesian network is an efficient classification model. However, so far some researchers still attempt to improve the performance by considering directions of edges, because traditional learning method merely takes into account log likelihood, which is not suitable for learning classifiers, when learning a tree topological structure. In this paper, we analyze the search spaces and the equivalent classes spaces of this kind of classifiers. Accordingly, we point out they are robust on structure learning because that the directions of their edges do not play a role in maximizing log conditional likelihood. For application, we propose a novel framework for learning these kind of classifiers. Finally, we run experiments on Weka platform using 45 problems from the University of California at Irvine repository. Experimental results show classification accuracy and stability do not change statistically in our learning framework.