An email modelling approach for neural network spam filtering to improve score-based anti-spam systems

This research proposes a model for presenting email to Artificial Neural Network (ANN) to classify spam and legitimate emails. The proposed model based on selecting wise 13 fixed features relevant to spam emails combined with text features. The experiment tests many scenarios to find out the best-suited combination of features representation. These scenarios show the effect of using term frequency (tf), term frequency-inverse document frequency (tf*idf), Level two (L2) normalization, and principal component analysis (PCA) for dimension reduction. Text features vectors are represented in the principal component space as a reduced form of the original features vectors. PCA reduction effect on ANN performance is also studied. Among these tests, best-suited model that improves ANN classification and speeds up training is concluded and suggested. An idea of integrating ANN anti-spam filter into score-based anti-spam systems is also explained in this paper. XEAMS email gateway, the commercial anti-spam, already uses Naïve Bayes (NB) filter as one of its many techniques to identify spam email. The proposed approach influences filtering results by 7.5% closer to XEAMS anti-spam system results than NB filter does on real-life emails of Arabic and English messages.