Naive Bayes spam filtering using word-position-based attributes and length-sensitive classification thresholds

This paper explores the use of the naive Bayes classifier as the basis for personalised spam filters. Several machine learning algorithms, including variants of naive Bayes, have previously been used for this purpose, but the author’s implementation using word-position-based attribute vectors gave very good results when tested on several publicly available corpora. The effects of various forms of attribute selection – removal of frequent and infrequent words, respectively, and by using mutual information – are investigated. It is also shown how n-grams, with n > 1, may be used to boost classification performance. Finally, an efficient weighting scheme for costsensitive classification is introduced.