Research On User Churn Prediction Technology Based On Hybrid Model In Telecom Field

In the field of telecommunications,user churn prediction refers to operators predicting the user to be churned before the user is churn,so that it can continue to use the services provided by operators to create profits.User churn prediction can help operators reduce the user churn,which is important for companies to increase revenue and improve competitiveness.However,due to problems such as the sparsity and imbalance of data in the field of telecommunications,the prediction of user churn in telecom field at home and abroad is mostly in the research stage,and there are few practical applications.In this paper,the telecom operator dataset and KDD Cup game dataset are used,both of which have data characteristics in the field of telecommunications.Data sparseness refers to the existence of a large number of null values in the dataset.In this paper,the null value is filled by adding 0 values or averages.The off-network rate of the telecom operator dataset used in this paper is about 1.28%,and the off-network rate of KDD Cup competition data is about 7.34%,and there is a data imbalance,which will seriously affect the predicted effect.This paper uses the machine learning combined with naive random oversampling to solve the problem of data imbalance,experimental results show that the AUC value reached 0.71602 and 0.68574 on telecom operator dataset and the KDD Cup dataset after using naive random oversampling.To maintain users,operators need to consider maintaining costs,and the way to reduce costs is using different maintenance solutions for users with different needs.In this paper,the idea of integration is used to propose two model mixes.The one-stage model mix is mainly combined with GDBT,AdaBoost and XGBoost models using Bagging method,which improves the prediction accuracy.After one-stage model blend,the AUC value of the telecom operator dataset and the KDD Cup dataset was increased to 0.71987 and 0.69571.The two-stage model mix mainly uses the GDBT and AdaBoost mixed GDBT_ADA,LR and XGBoost,two-stage of the quadratic model mix is to find high-risk users.Experimental results show that the combination of naive random oversampling and two stages models can effectively improve the accuracy and usability of the model.