Email Classification Using Deep Learning

Abstract

With the advent of business intelligence and big data, the machines have started sharing the workload of humans in various domains of work. In this study the task of email classification has been addressed through deep learning algorithms. The proposed system contributes to Startford university’s project of automatic reply agent in order to automatically respond to the queries of students without the intervention of human. To achieve the said objective three deep learning algorithms (Multilayer Perceptron, Recurrent Neural Network and Convolutional Neural Network) have been applied. The proposed system has been trained and tested with 3 datasets of varying length, domain and diversity respectively. Among the datasets, two are publicly available, i.e. ENRON Intent and University Helpdesk however, the third dataset AskTeacher2018 is generated by real time queries of students from the universities of Pakistan. The designed algorithms have been optimized in two stages; basic network optimization (batch size, epochs and no. of neurons etc.) and Advanced Network Optimization (Variation of the models, Non linearity, batch normalization etc.). Over and above, other techniques of testing and pre-processing have also been applied in order to improve the accuracy of the results and impact of all these factors have been observed and analyzed. The finest accuracy achieved for the three algorithms are 93.16% for ENRON Intent, 96.88% for University Helpdesk and 98.86% for Ask Teacher2018. Keywords: Deep Learning, Email Classification, Neural Network