Combining active learning and boosting for Naïve Bayes text classifiers

This paper presents a variant of the AdaBoost algorithm for boosting Naïve Bayes text classifier, called AdaBUS, which combines active learning with boosting algorithm. Boosting has been evaluated to effectively improve the accuracy of machine-learning based classifiers. However, Naïve Bayes classifier, which is remarkably successful in practice for text classification problems, is known not to work well with the boosting technique due to its instability of base classifiers. The proposed algorithm focuses on boosting Naïve Bayes classifiers by performing active learning at each iteration of boosting process. The basic idea is to induce perturbation of base classifiers by augmenting the training set with the most informative unlabeled documents.