The Study And Realization Of Customer-churn Model Based On Data Mining In Telcom

In recent years, customer churn in telecom becomes serious. For the three operatorsincluding Mobile,China Unicom and Telcom who can assure their customers not to abandon thedefault go to another one, at the same time can get the lost from others, will be the final winner.A bitter contest has launched among operators who do all they could to gain more customers.However, operators pay much more attention to customers churned rather than those who arereducing their consumption. Hardly realize that those people are running off gradually aspotential loss customers.Telecommunications industry has a mass of data and diversity which means that everycustomer has a large number of attributes which called variables in the data mining model, suchas ARPU, charge way, downtimes etc. In order to better model, this paper designs wide sheetfrom three aspects of operator sends messages to remind customer,customer perceptive value andvalue behavior of customer based on the assumption that the reasons for customers churn. Thendivides these variables into different groups to decide which variables to participate in modelmodeling. Combined the customer’s own characteristics and the grouping variables thensubdivide all of the customers.Customer-churn model becomes more and more in telecommunications industry, in order to improve thehit rate of the model, this paper proposes a combined model philosophy. The combined model is based on theconstraint model, prediction model, mark model. Constraint model selects variables has large distinction to beconstraint conditions, prediction model screens visible loss and relatively obvious variables,mark modelselects implicit customer churn variables to make up small amount of samples to identify customer morecomprehensively. Each model has its own special variables,as a result,the combination model proposed in thispaper plays a significant role in customer-churn model.This model uses IBM SPSS Statistics, puts forward the concept of pheromone difference inant colony algorithm, using the improved ant colony algorithm to subdivide the customer toimprovethe customer clustering effect.Canonical transformation eliminate the impact of thedimensionless coefficients to make the model more regular. Using logistic regression algorithminstead of decision tree to avoid the problem of little distinction of variables.Increasing logistic regression algorithm iterations to make the model more precise.In addition,analysis the variablesby the decision tree algorithm and factor analysis algorithm on depend.The final model evaluation shows that the combination model CPM is dynamic, stable andcomprehensive. Constraint model reflects dynamic changes in business, prediction modelguarantees a stable and high precision model using stable dominance variables, mark modelmakes up the small amount of sample to identify customer sample more carefulcomprehensively.