Recursive Autonomy Identification-Based Learning of Augmented Naive Bayes Classifiers

Abstract

Earlier reports have described classification accuracies of exactly learned augmented naive Bayes (ANB) classifiers. Those results indicate that a class variable with no parent has higher accuracy than those of other Bayesian network classifiers. Additionally, asymptotic estimation of the class posterior identical to that of the exactly learned Bayesian network is guaranteed to be achieved. Nevertheless, exact learning of large ANB is difficult because it entails an associated NP-hard problem that worsens as the number of variables increases. Recent reports have described that constraint-based learning methods with Bayes factor achieve larger network structures than when using traditional methods. This study proposes an efficient learning algorithm of an ANB classifier using recursive autonomy identification (RAI) with Bayes factor. A unique benefit of the proposed method is that the proposed method is guaranteed to accelerate execution of the RAI algorithm when the data follow an ANB structure. Numerical experiments were conducted to demonstrate the effectiveness of the proposed method.