Recurrent Neural Networks for Spam E-Mail Classification on an Agglutinative Language

Abstract

In this study, we have provided an alternative solution to spam and legitimate email classification problem. The different deep learning architectures are applied on two feature selection methods, including the Mutual Information (MI) and Weighted Mutual Information (WMI). Firstly, feature selection methods including WMI and MI are applied to reduce number of selected terms. Secondly, the feature vectors are constructed with concept of the bag-of-words (BoW) model. Finally, the performance of system is analyzed with using Artificial Neural Network (ANN), Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BILSTM) models. After experimental simulations, we have observed that there is a competition between detection results of using WMI and MI when commented with accuracy rates for the agglutinative language, namely Turkish. The experimental scores show that the LSTM and BILSTM give 100% accuracy scores when combined with MI or WMI, for spam and legitimate emails. However, for particular cross validation, the performance WMI is higher than MI features in terms e-mail grouping. It turns out that WMI and MI with deep learning architectures seem more robust to spam email detection when considering the high detection scores.