Analysis of machine learning-based spam filtering techniques

In this thesis, automatic spam e-mail detection problem is examined. Some existing machine learning algorithms are tested on an open dataset and the results are analyzed. The methods we developed have been implemented using machine learning and text classification techniques. We have used different data sets to develop and test the methods. The proposed methods for solving the problem are based on using weighted TF-IDF, SciKit Learn and Word2Vec vectorization. We developed and used vector representation methods for email text and then used supervised machine learning algorithms to classify emails as spam or ham. We used WEKA software tool to apply machine learning classification methods on vector representations of email. For classifications, we used the algorithms Support Vector Mechanism SVM (POLY), SVM (RBF), Naive Bayes, Bayesian Networks, J48 and Random Forest algorithms. We compared and analyzed the results we obtained from the classification methods. Our results show that the Word2Vec vector and the SVM (poly) algorithm perform better with 98.33% spam detection accuracy for 300 email data set.