Loss Prediction Based On Svm Customers

Client churn problems are currently prevailing in the mobile communications industry and also are one of the most concerns, clients churn prediction is the premise of this problem. This paper based on the platform of Yunnan mobile business analysis system, combined Yunnan mobile's operation requirement, analyzed related data mining theory and clients churn prediction arithmetic, distill Yunnan mobile clients churn data as sample data, adopt data mining process model, use support vector machine and Neural Network.etc arithmetic build clients churn prediction models. Through compare theory analysis and experiment results, proved the clients churn prediction model that adopt based on imbalanced core vector machine more suit to deal with mobile clients churn prediction problem.The modeling process of clients prediction model include:operation question definition, data selection, data prepare, data analysis, the build of model, results compare analysis and so on. But the model publish and implement comprise, because they are not the important parts of this paper, so we do not involved.First, this paper analyze the necessity and meaning of currently clients churn prediction, from a theoretical point analyze and compare the advantages and disadvantages of commonly used prediction models operation, and based on operation analysis system, analyze the features of mobile client churn problem. On the basis of statistical learning theory, analyze support vector machine basic theory and classification theory, apply non-equilibrium nuclear vector machine less used in currently customers churn prediction field to customers churn prediction problem. Secondly, accord to customers churn problem relevant definition, combine domain engineer experience and data mining knowledge to sample Yunnan mobile data,treat noise, analyze continuous variables and discrete variables of sampling data, decide samples set finally variables. In the experiment, point out experimental process and methods, establish LIBSVM and WEKA experimental platform, respectively generate non-equilibrium nuclear vector machine (ICSVM) model and neural network (BP), decision tree (C4.5), Bayes (NB), and introduce the required data format and call the method of different platform, explain the results treatment method. Finally in experimental results part, design the off-line rate, accuracy, training time and correct classification rate and so on, analyze ICSVM model other performance, from the perspective of theory and experiment proves that the model based on ICSVM fit deal with mobile clients churn prediction problem.