Measuring Performance of Spam Filter using Graham's Naive Bayes Classifier

Abstract

Spam email has become a major problem for Internet users and providers. After several failed attempt to filter spam based on heuristic approach such as black-listing or rule-based filtering, content-based filtering using naïve Bayes classifier has become the standard for spam filtering today. However, the naive Bayes classifier exists in different forms. This research aims to compare two different forms of naïve Bayes which are multinomial naive Bayes using boolean attribute and Graham version of naïve Bayes which is popular among several commercial and open source spam filter applications. This research also compares performace of two different methods for data trainings which are train-everything (TEFT) and train-on-error (TOE). Finally, this research attempts to identify several hard-to-classify emails. The evaluation result showed that multinomial naïve Bayes had better performance compared to Graham naïve Bayes. The result also showed that TEFT successfully outperforms TOE in term of accuracy.