The Factors that Influence Postpaid Subscribers to Make Churn in a Romanian Telecommunication Company

Andreea Dumitrache, Monica Mihaela Matei Maer, Stelian Stancu, Oana Popescu
2019 Proceedings of the International Conference on Applied Statistics  
The telecommunication industry is growing every day, increasing its competitiveness. In almost all European countries, the market penetration of mobile network users exceeded 100% (for example in Croatia it is over 130%). Acquiring new users is virtually impossible because there are no new users. There are only users of rival companies who are exposed to numerous marketing campaigns carefully designed to try to win them. That's why customer retention activity and churn prevention is a
more » ... tion is a necessity. The purpose of this paper is to predict customers who are willing to migrate to another Romanian mobile telecommunications company and to determine the strongest factors of influence in the consumer's decision to leave their current service provider for another provider. Migration behavior analysis is developed for customers with postpaid subscriptions. We applied the ROSE package for re-sampling and decision trees on the dataset to identify decision makers in the migration process. The combination of the two techniques in our study did not significantly improve the performance of the classifier measured by the AUC (Area Under the Curve). After balancing the sample, however, we obtain the optimal value of the AUC coefficient (0.724) for the second cluster, making the correct prediction of the churn phenomenon on the analyzed data set. The study is an addition of Churn Analysis in Romanian Telecommunications Company, M. M. Matei Maer and A. Dumitrache (2018), where ROSE and logistic regression was applied to the same dataset for the same purpose: balancing the sample and churn prediction, but the value of the AUC coefficient was really low, making it difficult to accurately predict the churn phenomenon. Therefore, another purpose of the current paper is to compare the performance of the two techniques used in combination with ROSE on the same set of data.
doi:10.2478/icas-2019-0021 fatcat:6xggzdpabbdzdc32t3kebyy55e