Lazy Associative Classification for Content-based Spam Detection

Despite all tricks and mechanisms spammers use to avoid detection, one fact is certain: spammers have to deliver their message, whatever it is. This fact makes the message itself a weak point of spammers, and thus special attention has being devoted to content-based spam detection. In this paper we introduce a novel pattern discovery approach for spam detection. The proposed approach discovers patterns hidden in the message, and then it builds a classification model by exploring the associations among the discovered patterns. The model is composed by rules, showing the relationships between the discovered patterns and classes (i.e., spam/legitimate message). Differently from typical eager classifiers which build a single model that is good on average for all messages, our lazy approach builds a specific model for each message being classified, possibly taking advantage of particular characteristics of the message. We evaluate our approach under the TREC 2005 Spam Track evaluation framework, in which a chronological sequence of messages is presented sequentially to the filter for classification, and the filter is continuously trained with incremental feedback. Our results indicate that the proposed approach can eliminate almost 99% of spam while incurring 0.4% legitimate email loss. Further, our approach is also efficient in terms of computational complexity, being able to classify more than one hundred messages per second