Comparative Study of Classification Approaches for Email Analysis

Abstract

Illicit messages like threatening and abusive messages affect emotions and psychology of a person. Such messages start exerting influence on mental status, and ultimately physical condition of a person. E-mails are one of the popularly used sources, for communicating personal and official messages. Typically, sentiment analysis of these e-mails includes classifying them into positive, negative and neutral messages. Identifying the sentiments of e-mails using an efficient and effective algorithm is very important and useful step in the domain of e-mail forensics. In this work, support vector machine, k-nearest neighbour, and neural network back-propagation algorithms are used to classify the sentiments of e-mail into positive, negative and neutral categories using self-curated e-mail dataset. This dataset is a combination of Enron's e-mail dataset and publicly available messages converted into e-mails. This paper presents a comparative study of classification approaches for e-mail analysis. Finally, it is concluded that the neural network with the back-propagation training algorithm provides the best results considering the accuracy and the memory requirements with the little compromise on the time required to recognise the sentiment of a given e-mail.