Multiple Machine Learning Algorithms for Spam Mail Detection

Abstract

Due to several issues including clogging mailboxes and obscuring critical personal correspondence caused by a large amount of spam mail, spam filtering application based on machine learning models has become more and more essential. This paper gives a comprehensive evaluation of 4 classic machine learning models including the Naïve Bayes classifier, support vector machine, neural network, and the decision tree in the task of spam email evaluation. In order to make fair comparisons between the algorithms, they were all built by learning from the same training dataset and asked to make predictions on the same testing dataset. More specifically, the Naïve Bayes classifier uses the conditional probability of the Naïve Bayes theorem and consumes the words that appear independently in the emails. The decision tree algorithm keeps recursively splitting the dataset to build a predictive theory. Support vector machine is able to make predictions by transforming the training dataset to a linear equation in n-dimensions. And neural networks can recognize the non-linear relationships between the inputs and outputs by constructing the front, hidden, and output layers. It is observed from the result that all machine learning models have successfully classified spam emails with high testing accuracy which are all above 90%. More specifically, the neural network achieves the highest testing accuracy followed by the support vector machine and decision tree. While the Naïve Bayes classifier does not perform well enough as it must assume the appearance of words happens independently.