Applying over 100 classifiers for churn prediction in telecom companies

Abstract

In today’s date where machine learning is the key to solve so many problems in different fields, one really should know the extent of its importance in their field. One of the major applications of machine learning is Predictive Analytics. Churn prediction is one of the key steps for customer retention in this saturating market scenario [31]. This is one of the major objectives and any toolkit which can give insights on this can be really beneficial for any service providing companies. Furthermore, one of the major problems that business analysts face during this procedure is to decide which classifier to select. In the continuously evolving field of machine learning where developers are constantly coming up with new machine learning algorithms, it is often difficult for the analysts to have knowledge about the varied options. In our work, we try to analyze and compare the performance of over 100 classifiers in churn prediction of a telecom company. We have used renowned classifiers from different families. This work can serve as the first step for any data scientist who wants to develop a churn prediction system for their application. Also, we try to explore efficient algorithms that will give a better result. Churn prediction is a mildly imbalanced set of the problem which degrade the performance of classifiers. The highest accuracy is given by the Regularized Random Forest classifier. Since the problem is imbalanced, we also consider the area under the Receiver Operating Characteristic (ROC) curve and the classifier Bagging Random Forest produces the best result in this scenario.