Attribute augmented and weighted naive Bayes

Numerous enhancements have been proposed to mitigate the attribute conditional independence assumption in naive Bayes (NB). However, almost all of them only focus on the original attribute space. Due to the complexity of real-world applications, we argue that the discriminative information provided by the original attribute space might be insufficient for classification. Thus, in this study, we expect to discover some latent attributes beyond the original attribute space and propose a novel two-stage model called attribute augmented and weighted naive Bayes (A2WNB). At the first stage, we build multiple random one-dependence estimators (RODEs). Then we use each built RODE to classify each training instance in turn and define the predicted class labels as its latent attributes. At last, we construct the augmented attributes by concatenating the latent attributes with the original attributes. At the second stage, to alleviate the attribute redundancy, we optimize the augmented attributes’ weights by maximizing the conditional log-likelihood (CLL) of the built model. Extensive experimental results show that A2WNB significantly outperforms NB and all the other existing state-of-the-art competitors.