A Classifier Learning Method Based on Tree-Augmented Naive Bayes

The structure of Tree-Augmented Naive Bayes (TAN) forces each attribute node to have a class node and a attribute node as parent, which results in poor classification accuracy without considering correlation between each attribute node and the class node. In order to improve the classification accuracy of TAN, firstly, the TAN structure is proposed that allows each attribute node to have no parent or only one attribute node as parent. Then, a learning method of building the tree-like Bayesian classifier using a decomposable scoring function is proposed. Finally, the low-order Conditional Independency (CI) test is applied to eliminating the useless attribute, and then based on improved Bayesian Information Criterion (BIC) function, the classification model with acquired the parent node of each attribute node is established using the greedy algorithm. Through comprehensive experiments, the proposed classifier outperforms Naive Bayes (NB) and TAN on multiple classification, and the results prove that this learning method has certain advantages.