Improving Accuracy and Performance of Customer Churn Prediction Using Feature Reduction Algorithms

Abstract

Prediction of customer churn is one of the most essential activities in Customer Relationship Management (CRM). However, the state-of-the-art of the customer churn prediction approach only focuses on the classifier selection in improving the accuracy and performance of churn prediction, but rarely contemplate the feature reduction algorithms. Furthermore, there are numerous attributes that contribute to customer churn and it is crucial to determine the most substantial features in order to acquire the highest prediction accuracy and to improve the prediction performance. Feature reduction decreases the dimensionality of the information and may allow learning algorithms to function faster and more effectively and able to produce predictive models that deliver the highest rate of accuracy. In this research, we investigated and proposed two (2) different feature reduction algorithms which are Correlation based Feature Selection (CFS) and Information Gain (IG) and built classification models based on three 3) different classifiers, namely Bayes Net, Simple Logistic and Decision Table. Experimental results demonstrate that the performance of classifiers improves with the application of features reduction of the customer churn data set. A CFS feature reduction algorithm with the Decision Table classifier yields the highest accuracy of 92.08% and has the lowest RMSE of 0.2554. This study recommends the use of feature reduction algorithms in the context of CRM for churn prediction to improve accuracy and performance of customer churn prediction.