Research On Customer Churn Prediction In Telecommunication Based On Attribute Selection Of Rough Set

Customer churn prediction is one of most main problems in customerrelationship management(CRM), which involving multidisciplinary, multifactor.How to construct a high performance model so as to improve the prediction accuracyis a hot topic of customer churn prediction. But there are various kinds of factorsinfluencing the classification performance of prediction model, especially in the dataprocessing period, whose result will impact the classification performance of theconstructed model. However, the existing researches on telecommunication customerchurn prediction lack of some scientific method, some still use artificial method.Although the progressing of construction method of prediction model can improveaccuracy, the interpretability of prediction model is poor, that is means it can’t playtoo much benefit in decision making. In this context, making some exploration andresearch on data preprocessing method and acquisition of decision rules is a newaspect of telecommunication customer churn prediction, which has important theorysignificance and practical value. Rough set is a mathematical tool to deal withuncertain and imprecise problems, which can be used to dig out information with it’sablity of attribute reduction (feature selection) and acquisition of decision rules. Inthis paper, we focus on method of feature selection and acquisition of decision rulesbase on the extended model of rough set.First of all, the fuzzy rough set theory is used to deal with mixed data, thecomputing method of significance of attribute importance is improved, improvedCEBARKNC feature selection algorithm is designed, the experimental data is usedto feature selection, whose result is compared and analyzed with Hu qinghua’smethod.Secondly, the extended multigranulation rough set is studied, themultigranulation rough set is extended to fuzzy rough set to construct a βmulti-granulation fuzzy rough set, and the granular space reduction algorithm isdesigned. After that the experimental data is used to granular space selection and theresult is compared with the result under singlegranulation rough set condition. Theacquisition method of decision rules based on βmulti-granulation fuzzy rough set, and some decision rules made according to experimental data.Finally, a telecommunication enterprise’s data is used to feature selection withthe designed algorithms respectively, and the processed data is applied to constructprediction model. After that the decision rules are made so as to improve theinterpretability of the model.