Prediction Of Mobile Phone Customer Churn Based On BP-Cox Model

Since the establishment of the market,the research on customer churn prediction has been widely concerned for enterprises.Because there are many factors affecting customer churn,it is difficult to make effective predictions of customer churn.Now,there are a lot of methods for the customer churn prediction,among which,regression classification methods and neural network methods are used mostly.Regarding of the characteristics of the two methods,we propose a new BP neural network model based on Cox for the first time.Raw variables generate new variables after Cox regression.New variables adjust the weight of input variables in the model.In this paper,firstly,we introduce and discuss the two main methods of customer churn prediction: neural network and survival analysis method.In regression analysis method,through the empirical analysis,we found that Cox regression is superior to logistic regression and decision tree in customer churn prediction.In the neural network method,the empirical analysis shows that the standard BP neural network is better than the RBF neural network in convergence speed and prediction accuracy.Therefore,we combine the neural network and survival analysis and propose a new model,the BP-Cox model.However,the standard BP neural network has a slow convergence rate.And it is easy to fall into local minimum values.To overcome these shortcomings and improve the accuracy of prediction,this paper introduces BPFM,BPSAM,BPAM-I and BPAM-II momentum term method.It is used to change the weight of the standard BP neural network.Based on the introduction of the momentum term,the paper also introduces the average variable step size method,the adaptive step size method,the variable step size LMS algorithm and the improved variable step size method.It is used to change the learning rate of the standard BP neural network.Through the empirical,we found that the BPAM-II and the improved variable step size method are most effective when the weight update and the learning rate converge.BPAM-II and improved variable step size method are used to improve the standard BP neural network.Finally,compared with the standard BP neural network,Cox regression model and improved BP-Cox model,We found the improved BP-Cox model has the best prediction effect.