A two-stage feature selection method based on Fisher’s ratio and prediction risk for telecom customer churn prediction

Telecom customer churn prediction is crucial to the customer relationship management systems of telecom operators. It aims to predict a particular customer who is at a high risk of churning. The predicting process includes the steps of data pre-processing, imbalance processing, feature selection, classifier training and evaluation. A two-stage feature selection method based on fisher’s ratio and prediction risk was proposed, which took advantage of the filter feature selection method and wrapper feature selection method to solve the high dimensionality problem of telecom customer churn prediction. The method was evaluated on a real-world dataset, and the experimental results verify that it is able to reduce feature dimensionality and improve the performance of classifiers.