A grid search optimized extreme learning machine approach for customer churn prediction

Abstract

Customers' behaviors such as tendencies, loyalty status, satisfaction criteria show an alteration day by day due to the changing world. So, these behavior changes should be analyzed very well in every step of the decision-making process. Customer churn analysis is the determination of customers who tend to leave by analyzing the customer data with various methods before this situation occurs. This study aims to develop an Extreme Learning Machine based model for customer churn prediction problem and to determine the model parameters that provide the best performance. Grid search is used for hyperparameter tuning. Also modified accuracy calculation approach has been presented. The churn data set obtained from the UCI Machine Learning Repository has been used. Naive Bayes, k-Nearest Neighbor and Support Vector Machine methods are selected for performance comparison of the model. With a value of 93.1%, the best accuracy measure has been obtained with Extreme Learning Machine. Due to the low number of parameters to be determined and performance evaluation measures that compete with other models’ results, it can be said that the Extreme Learning Machine is highly effective and interesting in the solution of the problem.