A Double Layer Bayesian Classifier

Numerous approaches have been proposed to relax the conditional independence assumption of naive Bayes, the accuracy performance was indeed improved relative to naive Bayes when the assumption is violated. But most of the previous approaches treated the attribute relation in the same way for all class labels. In practice, this relation may be different for different class labels. This paper proposes a novel approach, by which the posterior probability of different class label is evaluated using different attribute relation. Experiment results indicate that the new approach obtains comparative performance relative to other modern Bayesian classifiers on some datasets, and on some other datasets it outperforms the others.