| As a very important part of customer relationship management, customer churn management is getting more and more attention. As an effective method of customer churn management, customer chum warning can reduce the unnecessary loss of customers by constructing early warning model, forecasting and analyzing the potential loss of customers, timely warning and taking the corresponding retention measures which can effectively reduce the customer chum rate. Communication operators have a large customer base, so they have a wealth of customer data. And they have a strong demand for customer chum warning management. In this context, this paper puts forward the research on the prediction model of communication customer chum based on data mining,combines the ability of data mining to extract effective information from massive data,and builds the model to carry out early warning research on the potential loss behavior of communication customers.After studying the research achievements of domestic and foreign scholars, this paper summarizes the construction of early warning model and the application of data mining algorithms in recent years. And introduces the theory of customer loss, data mining related theory and early warning model related technology, and lay the foundation of this paper.In the aspect of model data preparation, this paper takes the customer data of a city communication operation enterprise as the empirical object. This paper discusses and validates the useless feature deletion, missing value padding, data discretization, and unbalanced data equalization, which ensures the higher data quality of model construction.In the aspect of key feature selection, this paper analyzes the effect of the key feature selection methods, such as chi-square test, principal component analysis and Fisher ratio, in view of the characteristics of high characteristic dimension of communication customer data. It is found that different key feature selection methods have different effects on the loss warning model based on different algorithms. In comparison, the ability of Fisher ratio filtering to optimize feature subsets is stronger than chi-square test and principal component analysis. In this way, the prediction of the loss warning model based on different algorithms can get better prediction effect.In the aspect of early warning model construction, this paper proposes to construct communication customer churn combination early warning model. Compared with the general combination of early warning model, this paper adds a feature selection step based on Fisher ratio, and optimizes the training set according to the best feature subset of each individual early warning model. Based on three kinds of data mining algorithms,C5.0 decision tree, BP neural network and SVM are used to construct the basic communication customer churtn warning model. The weights of the optimal combination loss warning model are obtained by using the Lagrange function, which makes the deviation of the prediction result of the combined early warning model and the prediction result of each individual early warning model minimized. According to the prediction results of three single - line early - warning models,the combined loss warning model is constructed and the forecast results of the combined loss early warning model are obtained. The empirical results show that the combined loss early warning model is better than the forecast of the basic loss early warning model, which can reduce the income loss of the communication operation enterprises to a certain extent. |
