Evaluation of different SVM kernels for predicting customer churn

Churn prediction has been emerging as an essential research area in the commercial sector, specially in telecommunication business. Predicting future churners and retaining them through campaigning is a crucial job in this competitive telecommunication market. Although, several classical procedures exist for predicting churn, but recently Support Vector Machines (SVMs) are gaining popularity for providing excellent out-of-sample generalization. This paper aims to evaluate various SVM kernels to predict churners on an imbalanced distribution of churners and non-churners data set. The experiment was carried out on a telecommunication data taken by random sampling. Our study shows that linear kernel outperforms commonly used Radial Basis Function(RBF), sigmoid and polynomial kernels.