Classification of Spam Emails through Hierarchical Clustering and Supervised Learning

Spammers take advantage of email popularity to send indiscriminately unsolicited emails. Although researchers and organizations continuously develop anti-spam filters based on binary classification, spammers bypass them through new strategies, like word obfuscation or image-based spam. For the first time in literature, we propose to classify spam email in categories to improve the handle of already detected spam emails, instead of just using a binary model. First, we applied a hierarchical clustering algorithm to create SPEMC-11K (SPam EMail Classification), the first multi-class dataset, which contains three types of spam emails: Health and Technology, Personal Scams, and Sexual Content. Then, we used SPEMC-11K to evaluate the combination of TF-IDF and BOW encodings with Na"ive Bayes, Decision Trees and SVM classifiers. Finally, we recommend for the task of multi-class spam classification the use of (i) TF-IDF combined with SVM for the best micro F1 score performance, 95.39%, and (ii) TD-IDF along with NB for the fastest spam classification, analyzing an email in 2.13ms.