The Analysis On Warning Of Telecom Customer Churn Using Dada Mining

As one of the important business analysis topics in telecom operators’ eye, themain task of customer churn research is data mining to establish a prediction modeland to analyze behavioral characteristics base on the data of customers who havechurn already and those who are still using operator’s services, thus thedecision-maker can formulate marketing strategies. Using the CRISP-DM data miningprocess, this paper build a prediction model for mobile customer, which includesbusiness understanding, data understanding, data preparation, modeling, modelevaluation and model deployment. We defines customer churn firstly and thencompare seven classification algorithm often used in warning analysis, includinglogistic regression, decision tree, nearest neighbor classifier, Bayesian classifier,neural network, support vector machine model and classifiers combined. Based onthat, we select the customer behavior data of Tianjin mobile ltd. from October toNovember in2012, and then using the results of models to predict these customerwould whether churn or not. Comparing the empirical results of the seven modelsusing different algorithms, we have four findings:(1) the performance of decision treeand combined classifiers are the best without considering the problem of unbalanceclass. The prediction error rate of training data are respectively0.128%and0.235%,while the error rate of testing data are respectively4.84%and3.87%;(2) Thecomputing time of nearest neighbor classifier, neural network and support vectormachine model are much longer than other algorithms;（3）The prediction error rateof logistic regression model is very high although it can explain the basis howcustomer is classified;(4) The prediction error rate of Bayesian classifier is alsohigh because of the existing of correlation between these attribute variables selectedto describe customers’ behavior.