A Network-Based Model for Television Services Churn Prediction

Abstract

Predicting churn helps us understand which customers are likely to replace the company’s services with competitors. As the cost of acquiring users is much higher than retaining existing ones, churn prediction has emerged for numerous telecommunication companies as a critical tool to retain an existing customer base. Usually, churn is predicted by modeling individual customers’ behaviour and relatively static features such as demographic data, contractual data, and product information. Recent work has shown that analysing customers’ social network improves the accuracy of churn prediction. Although the network analysis is widely researched for telecommunication customers, little to no research was found for TV service users. This thesis attempts to fill this gap by analysing customers behaviour prior to churning as well as their call logs. Models with and without the network analysis features were trained with XGBoost, Adaboost, Random forest, Logistic regression, and Gradient Boost Classifier. Differences in the prediction results, whether the additional features were added, were presented in this paper. Results indicate that adding information from call logs improves the minority class prediction results.