Research And Application Of Data Mining In Telecom Customer Churn Prediction

With the issuance of 4G licenses at the end of 2013,the three telecom operators in China opened a more intense customer market competition.Many different marketing methods have been used to attract new customers,which will inevitably lead to the unstable state and the loss of large numbers of customers.Meanwhile,the customer churn can lead to the falling of operator's market share,which can directly cause economic losses to the operator.Therefore,the work on customer churn prediction has important practical significance.The China Unicom of some place has stored large amounts of customer data in their database,and needs to use these data to carry out the work on customer churn prediction.Combine with the actual situation,the main works of this study are as follows:1.The relevant researches at home and abroad on customer churn prediction were analyzed in detail from the following three aspects:the reason of customer churn,selection of prediction variables and models.2.By the analysis of real data,we found that the number of churned customers occupies only 7%of the total number of customers,and this indicated that there is serious unbalanced distribution of the data,which would have a great bad effect on the late modeling.To solve this problem,the under-sampling technique was used.The raw data was sampled with different ratio and then C5.0 decision tree,neural networks and Logistic regression were used to construct the prediction model on these dataset.3.Because each classification algorithm has its own advantages and disadvantages,we can't take full use of the information in data.For this problem,the ensemble of different classification algorithms was used to the churn prediction of telecom customer.C5.0 decision tree,neural networks and Logistic regression were chosen as the sub-classifiers.To calculate the ensemble weight coefficients,Lagrange equation based on sum of squared error was constructed.4.Ensemble prediction model for Unicom's customer churn prediction was constructed,and the potential churned customers were identified,which will provide a basis for the enterprise to make decision to retain potential churned customers.In summary,by using the data under-sampling technology,the prediction results of every classification algorithm have been improved,and the ratio of data under-sampling for different classification algorithms varies.Then,ensemble prediction method was used to the churn prediction of telecom customer,and the result is better than single prediction models.So applying the ensemble prediction method in customer churn prediction is meaningful.