Offering a hybrid approach of data mining to predict the customer churn based on bagging and boosting methods

Abstract

Purpose

Churn management is a fundamental process in firms to keep their customers. Therefore, predicting the customer’s churn is essential to facilitate such processes. The literature has introduced data mining approaches for this purpose. On the other hand, results indicate that performance of classification models increases by combining two or more techniques. The purpose of this paper is to propose a combined model based on clustering and ensemble classifiers.

Design/methodology/approach

Based on churn data set in Cell2Cell, single baseline classifiers, ensemble classifiers are used for comparisons. Specifically, self-organizing map (SOM) clustering technique, and four other classifier techniques including decision tree, artificial neural networks, support vector machine, and K-nearest neighbors were used. Moreover, for reduced dimensions of the features, principal component analysis (PCA) method was employed.

Findings

As results 14 models are compared with each other regarding accuracy, sensitivity, specification, F-measure, and AUC. The results showed that combination of SOM, PCA, and heterogeneous boosting achieved the best performance comparing with other classification models.

Originality/value

This study examined the performance of classifier ensembles in predicting customers churn. In particular, heterogeneous classifier ensembles such as bagging and boosting are compared.

Citation

Fathian, M., Hoseinpoor, Y. and Minaei-Bidgoli, B. (2016), "Offering a hybrid approach of data mining to predict the customer churn based on bagging and boosting methods", Kybernetes, Vol. 45 No. 5, pp. 732-743. https://doi.org/10.1108/K-07-2015-0172