Combined Bayesian classifiers applied to spam filtering problem

Abstract

This paper focuses on the problem of designing effective spam filters using combined Näive Bayes classifiers. Firstly, we describe different tokenization methods which allow us for extracting valuable features from the e-mails. The methods are used to create training sets for individual Bayesian classifiers, because different methods of feature extraction ensure the desirable diversity of classifier ensemble. Because of the lack of an adequate analytical methods of ensemble evaluation the most valuable and diverse committees are chosen on the basis of computer experiments which are carried out on the basis of our own spam dataset. Then the number of well known fusion methods using class labels and class supports are compared to establish the final proposition. © 2013 Springer-Verlag Berlin Heidelberg.