Forecasting Subscriber Churn: Comparison of Machine Learning Methods

Abstract

In order to remain competitive today in the telecommunications business, it is necessary to identify customers who are dissatisfied with the services provided. Therefore, forecasting subscriber churn has become an essential issue in this area. This article overviews different machine learning techniques including Decision Trees (DT), Naive Bayes Classifier (NB), Random Forest (RF), Artificial Neural Network (NN), KNearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM) and their ensembles (bagging and boosting) in order to demonstrate the superiority of the CatBoost technology in gaging the effectiveness of classifiers. To achieve the goal, data was classified and the specific advantages, when compared to others, of the CatBoost method were revealed based on obtained results. For the study, we analyzed four databases: 3 datasets are in open access and 1 dataset was provided by a Russian mobile company. Often, the dimension of these databases is high, which leads to a number of problems (including class imbalances, parameter correlations), which are solved by employing the dimensionality reduction method: Principal Component Analysis (PCA). The results obtained are compared with each other as well as with the results presented by other researchers based on open databases. The effectiveness of classifiers is evaluated using measures such as the area under the curve (AUC), accuracy, F1 -measure, and time.