Adaptive telecom churn prediction for concept-sensitive imbalance data streams

Abstract

A key toward intelligent decision-making in industries lies in the ability to process and analyze vast quantities of business data. Concept drift and class imbalance are co-existing problems in real-life data sets, and the telecommunication sector is no exception. It has been discovered only recently that the problems of concept drift and class imbalance, which were thought to be totally independent, are correlated in many ways and they have adverse effects on each other. Cumulative Sum Detection Method (CusumDM), a simple and efficient method based on Cumulative Sum Charts, handles the problem of concept drift efficiently. However, the presence of class imbalance causes a decline in its performance. To facilitate intelligent customer churn prediction, this article proposes optimized two-sided Cusum churn detector (OTCCD), which is a significant improvement of CusumDM that handles the problems of class imbalance and concept drift by determining error rate of sliding windows alongside. This approach has been applied to the Call Detail Record (CDR) of a South Asian Telecom Company for churn prediction. Handling telecom data requires high-performance computing power and resource-aware intelligent methods due to the velocity and speed of the data. Presence of an extremely low number of churners creates class imbalance problems in CDR and abruptly, or gradually changing data distribution makes the churn data a rich source to target these problems. Prediction accuracy and mean evaluation time were the basis for verification of classification results, showing that OTCCD outperformed its antecedents CusumDM, Ensemble Drift Detection Method (Ensemble) and Adaptive Windowing (ADWIN) Change Detector by producing better accuracy.