Explainable Machine Learning: A Manuscript on the Customer Churn in the Telecommunications Industry

The work in this manuscript is an application of an explainable machine learning approach to predicting customer churn for a fictitious telecommunications company. When developing a customer churn prediction model, it is critical to understand why each prediction is made in addition to achieving high accuracy in the model’s predictions. The purpose of this work is to demonstrate the utility of machine learning model explanation techniques by using a case study to analyze the results of various machine learning models (decision tree, random forest, and light gradient boosting machine) that predict individuals at risk of churning. The light gradient boosting machine achieved the highest AUC score of 0.87, as well as a recall score of 0.95 for churning customers. The decision tree and random forest misclassified churning customers with high probabilities of 28% and 19%, respectively, whereas the light gradient boosting machine had the lowest misclassification probability of 6%. Telecommunications service providers prioritize customer retention over new customer acquisition because it is less expensive to retain customers than to acquire new ones, so we want to have as few false negatives as possible. When compared to the other models, the light gradient boosting machine has the fewest false negatives (0.05%). The interpretation of the decision tree is given by a tree decomposition, showing rules that leads to a particular prediction (churn or non-churn). Local interpretable model-agnostic explanations and shapley additive explanations were used to explain the decision making of the complex models (random forest and light gradient boosting machine). A deeper insight into how the model behaves when predicting the customer churn is obtained.