Comparative Analysis of Machine Learning Techniques for Churn Prediction

Abstract:

Getting customers to come back has become an essential strategy across industry as old customers tend to be cheaper to keep, compared to attracting new customers. Thus a comparative analysis of machine learning techniques for customer churn prediction assume a great importance in order to enhance the customer relationship management and guide the companies in making the best possible retention strategies. This study evaluated the performance of six machine learning algorithms in accurately predicting customer churn. The algorithms employed include Logistic Regression, K-Nearest Neighbours, XGBoost, Random Forest, AdaBoost, and Gradient Boosting. The study employed multiple data balancing techniques such as SMOTE, SVMSMOTE, BorderlineSMOTE, ADASYN and Random Oversampling to imbalanced class and enhances the model performance using three banking related churn datasets. Each model was evaluated using accuracy, precision, recall, specificity, G-mean, ROC score and MCC. The results indicate that ensemble methods such as Random Forest and Gradient Boosting outperform other machine learning algorithms used, while data balancing techniques like SMOTE, BorderlineSMOTE, and SVMSMOTE improve recall and sensitivity, key factors in identifying ―at risk‖ customers.

Keywords: Churn, churn prediction, customer retention, ML, XGboost, Random forest. K-Nearest neighbours, Logistics regression