A Study Of Telecom Customer Churn Prediction Model Based On Cost-Sensitive Decision Tree

With the fiercer competition in telecom industry,the customer churn rate is increasing.So,to predict churn precisely becomes a key factor of the existence and development of telecom corporations.There are two common problems in telecom data set.The one is the distribution unbalance between churn set and non-churn set,the other one is the numerous missing values. To solve these problems,this paper brings forward the cost-sensitive learning theory.The misclassification cost and the test cost are brought into the decision tree building process,based on the traditional C4.5 algorithm.In order to eliminate the harmful effect caused by the unbalance of the telecom data set, the two different samples are assigned with different misclassification costs. Meanwhile,to reduce the loss resulted by the missing values,different attributes are assigned with different test costs.Besides,a redistribution method based on costs is generalized and the two cost's quantitative description is set up.A customer churn prediction model suiting for the present telecom data set is set up based above research.This technology combined with the telecom background has improved present cost-sensitive method and compensated the shortcoming of the traditional telecom churn prediction method.The relative experiments have shown that,the churn prediction model based on this new technology greatly enhanced the precision of the churn prediction,without reducing the precision of the total prediction.This is so meaningful for the churn characteristics grasping and the customer retaining strategy.