Some empirical results on two spam detection methods

In this paper, we describe the results of an empirical study on two spam detection methods: support vector machines (SVMs) and naive Bayes classifier (NBC). To conduct the study, we implement the NBC and choose to use the SVM/sup light/, an application of SVMs developed by Thorsten Joachims. The NBC and the linear SVMs with different C parameters are trained on a set of 2000 emails with 1000 spams and 1000 nonspams, and are tested on 200 new emails with 100 in each class. A program is written that converts emails into feature vectors using both TF and TF-IDF term weighting methods. The evaluation criteria include accuracy rate, recall, precision, miss rate, and false alarm rate. The results indicate that the both approaches have their pros and cons.