Evaluation of Three Discrete Methods on Customer Churn Model Based on Neural Network and Decision Tree in PHSS

Nowadays, churn prediction and management is critical for telecommunication companies in the fast changing and strongly competitive market. Our research objective was to compare the effectiveness of different discretization methods for predictor variables in building customer churn models of Personal Handy- phone System Service (PHSS), and to build an effective and accurate customer churn model of PHSS. Therefore, two experimentations including 24 churn models are put forward to compare and improve the prediction ability of churn models. Our research suggests that: (1) the more the number of categories of predictor variables is, the better the predictive ability of Model with different discretization methods is. (2) Predictive stability of models trained is very satisfying. (3) The method presented is effective and feasible under the condition that information is very little and class distribution is skewed.