Application Of Cost-sensitive XGBoost Algorithm In Telecom Subscriber Loss Prediction

With the popularization of mobile Internet,the communication customer market is gradually saturated,and the growth of telecom customers continues to slow down.While the operators are discovering new customers,the market competition among each other is intensifying.In recent years,the implementation of "Number portability" policy has accelerated the flow of customers in telecom enterprises.How to timely identify the customers with off-network tendency and launch personalized marketing strategies has become an urgent goal of the telecom industry.In the field of customer churn,there is a general problem of unbalanced categories,and the proportion of lost customers is small.The error cost of the traditional binary classification algorithm is the same,but the loss caused by classifying a lost customer as a non-lost customer is much greater than the loss caused by classifying a non-lost customer as a lost customer.Based on this,this paper introduces the idea of cost-sensitive learning,and sets different penalties for different classification errors.Furthermore,the newly developed integrated learning algorithm XGBoost is adopted as a customer churn forecast tool to obtain better forecast performance.The main work of this paper is to combine cost-sensitive learning with XGBoost algorithm by modifying the loss function,and to build a forecast model based on XGBoost.Based on the empirical analysis of the service data of mobile operators in a certain city,the results show that the XGBoost model with modified loss function performs better in all kinds of evaluation indexes than the model trained on default loss function.At the same time,the experiment comparing with the data sampling method also proves that for the data set in this paper,the cost-sensitive learning performs better in dealing with the unbalanced problem.It can be seen that the forecast model trained by the cost-sensitive XGBoost algorithm has a good classification effect on loss and normal samples,which can provide a solution for operators to timely find customers with off-network tendency and formulate corresponding maintenance strategies.