| With fierce competition in telecommunication, saturated capability of the market, relatively stable price of the terminal products and telecom service, customers are the focus of every telecommunication company .How to efficiently prevent customers from churning and reduce churn rate is becoming more and more urgent. High churn rate has caused huge lose for telecommunication companies. But it costs less to retain a existing customer than to develop a new one. Hence it's necessary to build a customers churn prediction model in telecommunication. It can effectively predict who will churn and find out the reasons of churn, so that the company can use some measures to retain the churning customers in advance.The purpose of this paper is to research and implementation a customer churn prediction model in telecommunication, it must have better accuracy and effectiveness. Theories of data mining and relative arithmetic are introduced .Then on the basis of a actual project, the design and implementation of prediction system are realized according to the CRISP_DM (Cross-industry Standard Process for Data Mining) framework. The sequence of demonstration is business understanding, data understanding, data preparation, modeling, evaluation and development.In this study, to address the highly skewed class distribution between churns and no-churns, the multi-classifier class-combiner approach was adopted. Firstly, multiple training subsets with a desired class ratio were created. Subsequently, for each training subset a classification model was generated by C5.0 technique. Then an overall prediction for the unseen instance was delivered by a meta-classifier using a weighted voting-based strategy.With this approach, the system does not result in a 'null' prediction system that simply predicts all instances as non-churners. It also can correctly predict more churners with a better accuracy rate than any other base classifier and improve the generalization of the model. |
PDF Full Text Request