Title: Sampling method based on improved C4.5 decision tree and its application in prediction of telecom customer churn

Authors: Weibin Deng; Linsen Deng; Jin Liu; Jie Qi

Addresses: School of Management and Economics, Chongqing University of Posts and Telecommunications, Nan'an district, Chongqing, China ' School of Management and Economics, Chongqing University of Posts and Telecommunications, Nan'an district, Chongqing, China ' School of Management and Economics, Chongqing University of Posts and Telecommunications, Nan'an district, Chongqing, China ' China Telecom Co., Ltd. Hefei Branch, 255, Changjiang West Road, Hefei, China

Abstract: Nowadays, customer churn prediction is quite important for telecom operators to reduce churn rate and remain competitive. However, the imbalance between the retained customers and the churners affects the prediction accuracy. For solving this problem, a new sampling method based on improved C4.5 decision tree is proposed. Firstly, an initial weight is set for each sample according to the data scale of each class. Then, the samples' weight is adjusted through several rounds of alternative training by the improved C4.5 decision tree algorithm. Both the gain ratio and the misclassification cost are considered for splitting criterion. Besides, the boundary minority examples and the centre majority examples are found according to their weights. Furthermore, over-sampling is conducted for the boundary minority examples by synthetic minority over-sampling technique (SMOTE) and under-sampling is executed for the majority examples. Experiments on UCI public data and telecom operator data show the efficiency of the new method.

Keywords: telecom customer churn; imbalanced data; under-sampling; over-sampling; decision tree; data mining.

DOI: 10.1504/IJITM.2019.097887

International Journal of Information Technology and Management, 2019 Vol.18 No.1, pp.93 - 109

Received: 24 Mar 2017
Accepted: 05 Feb 2018
Published online: 22 Feb 2019 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article