A 2-means Clustering Technique for Unsupervised Spam Filtering

Abstract

Unsolicited commercial e-mail, or “Spam”, implies a waste of network bandwidth and waste of human effort in internet and mobile phones communication. It is also a hard problem to distinguish legitimate from spam emails. The majority of the proposed algorithms use supervised learning techniques. Unfortunately, these approaches have the drawback of training over a large amount of manually and costly tagged email corpora. In this paper, we present an unsupervised method to address the problem of filtering spam emails without the need of training over such corpora. Using a 2-means clustering technique we perform a 2-way classification. To overcome the serious complications imposed by the large dimensionality of the data, the algorithm first transforms the data into a low dimensional component space applying a Principal Component Analysis over the data and then performs clustering on them.  The method was proved to be promising when evaluated over the publicly available corpus, called “SpamAssasin”, which is provided by the Open Project for evaluation purposes. The achieved performance is comparable to the performance of systems based on supervised learning techniques.