A Research On The Application Of Telecom Customer Churn Prediction Based On Random Forest

Customer churn can bring great loss to telecom companies,according to scientific research,the cost of developing a new customer is 6 times larger than that of retaining a frequent customer.Moreover,the profit brought by highly valued frequent customer is 16 times larger than that of a new customer,thus making it is vital to reduce customer churn.This paper focuses on using data mining technology to build a tow-class classifier to predict churn.The main findings and contributions are as follows:1.To tackle the high dimension dataset in telecom customer churn prediction,fisher score method is applied to select features and reduce dimension.By conducting experiments,decision tree,random forest and bagging models are trained on datasets with different number of features.2.Random oversampling(ROS),SMOTE and ADASYN are applied to get three different datasets to train decision tree,random forest and bagging models.The results show that ADASYN took a long time to oversampling datasets,which took over 10 times longer than that of SMOTE and ROS,and the performances of ADASYN on the three models are lower than SMOTE and ROS.ROS is the best method to oversampling datasets.3.Max profit(MP)criterion is applied based on cost-profit prospective to combine cost and profit when launching a retain campaign.4.A new proposed example-based cost-sensitive learning method is applied.By setting the cost of true positive,true negative,false positive and false negative of each example and adding these cost to feature selection criterion and pruning strategy,a decision tree-based random forest model is trained.The results show that using the proposed model,recall rate can increase to 97%,which is the best among bagging and random forest models.