Data mining via rules extracted from GMDH: An application to predict churn in bank credit cards

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. © 2010 Springer-Verlag.