Collaborative spam filtering based on incremental ontology learning

Abstract

Spam mail filtering is a classic problem to automatically recognize irrelevance between incoming emails and user contexts. This paper proposes a novel proxy server architecture for (i) collaboratively integrating useful features sent from personal email clients. (ii) Improving the filtering performance of SMTP servers. Given a set of spam mails marked by multiple email users, the proxy server can extract two kinds of textual features, which are apriori terms and concept terms based on key phrases. More importantly, by taking into account the semantics and statistical associations, the proxy can aggregate them in a hierarchical cluster structure. As a result, spam ontology can be built, and also, incrementally enriched. Hence, the email clients can be supported to improve their performances of spam filtering by referring to the semantic information from the ontology. For evaluating the proposed system, we have collected a large number of spam mails within a same intranet environment. The system has shown 17.4% lower error rate of filtering than the single email clients.