Marketing research suggests that it is more expensive to recruit a new customer than to retain an existing customer. In order to retain existing customers, academics and practitioners have developed churn prediction models to effectively manage customer churn. In this paper, we propose two genetic-algorithm (GA) based neural network (NN) models to predict customer churn in subscription of wireless services. Our first GA based NN model uses a cross entropy based criterion to predict customer churn, and our second GA based NN model attempts to directly maximize the prediction accuracy of customer churn. Using real-world cellular wireless services dataset and three different sizes of NNs, we compare the two GA based NN models with a statistical z-score model using several model evaluation criteria, which include prediction accuracy, top 10% decile lift and area under receiver operating characteristics (ROC) curve. The results of our experiments indicate that both GA based NN models outperform the statistical z-score model on all performance criteria. Further, we observe that medium sized NNs perform best and the cross entropy based criterion may be more resistant to overfitting outliers in training dataset.
All Science Journal Classification (ASJC) codes
- Computer Science Applications
- Artificial Intelligence