Spam E-Mail Classification by Utilizing N-Gram Features of Hyperlink Texts

With the advent of the Internet and reduction of the costs in digital communication, spam has become a key problem in several types of media (i.e. email, social media and micro blog). Further, in recent years, email spamming in particular has been subjected to an exponentially growing threat which affects both individuals and business world. Hence, a large number of studies have been proposed in order to combat with spam emails. In this study, instead of subject or body components of emails, pure use of hyperlink texts along with word level n-gram indexing schema is proposed for the first time in order to generate features to be employed in a spam/ham email classifier. Since the length of link texts in e-mails does not exceed sentence level, we have limited the n-gram indexing up to trigram schema. Throughout the study, provided by COMODO Inc, a novel large scale dataset covering 50.000 link texts belonging to spam and ham emails has been used for feature extraction and performance evaluation. In order to generate the required vocabularies; unigrams, bigrams and trigrams models have been generated. Next, including one active learner, three different machine learning methods (Support Vector Machines, SVM-Pegasos and Naive Bayes) have been employed to classify each link. According to the results of the experiments, classification using trigram based bag-of-words representation reaches up to 98,75% accuracy which outperforms unigram and bigram schemas. Apart from having high accuracy, the proposed approach also preserves privacy of the customers since it does not require any kind of analysis on body contents of e-mails.