Learning augmented Bayesian classifiers: A comparison of distribution-based and classification-based approaches

The naïve Bayes classifier is built on the assumption of conditional independence between the attributes given the class. The algorithm has been shown to be surprisingly robust to obvious violations of this condition, but it is natural to ask if it is possible to further improve the accuracy by relaxing this assumption. We examine an approach where naïve Bayes is augmented by the addition of correlation arcs between attributes. We explore two methods for finding the set of augmenting arcs, a greedy hillclimbing search, and a novel, more computationally efficient algorithm that we call SuperParent. We compare these methods to TAN; a state-of the-art distribution-based approach to finding the augmenting arcs.