Machine learning techniques applied to prepaid subscribers: Case study on the telecom industry of Morocco

In order to survive the fierce competition in today's telecommunication industry, it is mandatory to understand the need of customers who might think to move toward another competitor. Thus, assessing the churn prediction, which becomes a real concern in the telecom industry, is critical in predicting future trends of the industry. In this work, we wanted to determine the best and reliable prediction model in improving the competitiveness of Moroccan telecommunication sector. We focused on prepaid costumers, a group of subscribers that are unreliable, less committed to a provider, and they can end their services without prior notices. We used various machine learning techniques applied to Call Details Records (CDR) to analyze a dataset of 635997 subscribers and monitored their usage behaviors over a period of three months. The Visitor Location Register ID, which describes the locations where calls/SMS/Recharge and Data events were performed was also used to investigate the churning rate per different regions. Among models tested, we found that the Gradient Boosting Classifier (GBC) is the best classifier model with accuracy of 91% and f1 score of 89%. This model fits better the Churn prediction use case.