Research And Application Of Teleconmunication Operator User Churn Prediction Based On Data Mining

Users are the key to operator interests.In face of fierce competition,the increasingly saturated telecommunications market and the constant impact of the Internet tide,predicting users churn in advance and retaining users in a targeted manner is important to improve operator's profits.Data mining helps to obtain valuable information from data.Using data mining technology to model of massive data can accurately predict users churn.Based on data mining,this dissertation analyzes the data in the operator billing system and constructs the operator churn prediction system,which makes the user retention strategy more targeted and effective.It has a guiding significance for the operator's retention strategy formulation.The main work of the dissertation includes:Firstly,the data is preprocessed,characterized and selected.With LightGBM as the classifier,the operator user churn prediction model is proposed and tested.Results show all the correct rate,recall rate,F1 value and AUC of the proposed model are better than logistic regression,support vector machine,decision tree and random forest model.Also this dissertation builds a user churn prediction system that is easy to deploy and apply.It uses the real data extracted by the operator billing system to perform system testing.Secondly,this dissertation studies the imbalance problem from two aspects of data processing and classification.In terms of data processing,this dissertation is aimed to construct the churn user data based on generative adversarial networks,which improves the churn user identification of LightGBM.The effects of this method and random oversampling,SMOTE,Borderline-SMOTE and ADASYN on unbalanced user churn prediction are compared.Results show that the proposed method has a better effect in improving the correct rate,recall rate and F1 value.Thirdly,in terms of classification,the EasyEnsemble algorithm is improved.Both of the SMOTE algorithm and the ENN algorithm are merged into the EasyEnsemble framework to improve the classification when data is extremely unbalanced.The operator user churn prediction model based on the improved EasyEnsemble algorithm is constructed.The performance test results show that this algorithm has better recall effect and higher F1 value.On this basis,this dissertation combines the above two methods to further improve the performance of the model.