Data Mining via Rules Extracted from GMDH: An Application to Predict Churn in Bank Credit Cards

Abstract

Abstract. This paper proposes a hybrid method to extract rules from the trained Group Method of Data Handling (GMDH) neural network using Decision Tree (DT). The outputs predicted by the GMDH for the training set along with the input variables are fed to the DT for extracting the rules. The effectiveness of the proposed hybrid is evaluated on four benchmark datasets namely Iris, Wine, US Congressional, New Thyroid and one small scale data mining dataset churn prediction using 10-fold cross-validation. One important conclusion from the study is that we obtained statistically significant accuracies at 1% level in the case of churn prediction and IRIS datasets. Further, in the present study, we noticed that the rule base size of proposed hybrid is less in churn prediction and IRIS datasets when compared to that of the DT and equal in the case of remaining datasets.

A diffusion model for churn prediction based on sociometric theory

“…The monthly churn rates can reach and even exceed 15 %, depending on the considered service provider and the market this service provider operates in. Although churn is an important revenue factor in the telecommunication business, other businesses, such as food-based retailing (Miguéis et al 2012), the credit-card business (Naveen et al 2010), online gaming (Kawale et al 2009), and advertising (Yoon et al 2010), also face this problem.…”