Predicting email opens using machine learning techniques

Abstract

In this thesis we use machine learning techniques to predict email campaign open rates based on the subject line. We tackle the following two specific problems: “Given the subject line, what will be the open rate of the email for a target audience?” and “Given a subject line, will a specific user open the email?”. Both problems are solved in the context of one company that has a small email campaign dataset combined with a large audience. For the first problem, we use regression models such as Ridge regression to predict the open rate of a single subject line for the target audience of the company. We analyze the effect of multiple pre-processors such as scales and Principal Component Analysis. However, due to the lack of email subject line data this model is outperformed by a simple statistical model that utilizes the individual user data. Next, for user-specific open predictions, there is no baseline or simple statistical model. Instead, we propose a method that uses classification models, such as logistic regression and a random forest classifier, to predict for each individual user if they will open an email with a specific subject line by combining the features from the user and an email. We will demonstrate that word embeddings have a significant influence on the machine learning model and that the model performance can improve by only using emails form a single language. We have devised a method of calculating the open rate of an email from the user-specific predictions.