Words versus character N-grams for anti-spam filtering

The increasing number of unsolicited e-mail messages (spam) reveals the need for the development of reliable anti-spam filters. The vast majority of content-based techniques rely on word-based representation of messages. Such approaches require reliable tokenizers for detecting the token boundaries. As a consequence, a common practice of spammers is to attempt to confuse tokenizers using unexpected punctuation marks or special characters within the message. In this paper we explore an alternative low-level representation based on character n-grams which avoids the use of tokcnizers and other language-dependent tools. Based on experiments on two well-known benchmark corpora and a variety of evaluation measures, we show that character n-grams are more reliable features than word-tokens despite the fact that they increase the dimensionality of the problem. Moreover, we propose a method for extracting variable-length n-grams which produces optimal classifiers among the examined models under cost-sensitive evaluation.