Churn prediction with Linear Discriminant Boosting algorithm

Abstract

In this paper, a novel classification algorithm called Linear Discriminant Boosting (LD-Boosting) is proposed. By aggregating LDA learning through the boosting framework, this algorithm can deal with complicated binary classification problems, especially problems such as churn prediction with extremely imbalanced dataset. LD-Boosting is efficient since the most discriminative feature is computed in closed form in each iteration, with neither time-consuming numerical optimization nor exhaustive search. Furthermore, because of the computational simplicity of LDA learning, the method is able to utilize huge amount of training samples efficiently. In addition, boosting technique is employed in this algorithm to put heavier penalties on misclassification of the minority class, therefore directly reduces error cases and achieves more precise prediction results. The effectiveness of the proposed algorithm is validated by churn prediction experiments on a real bank customer churn data set. The method is found to improve prediction accuracy significantly compared with other algorithms, such as artificial neural networks, decision trees, support vector machines, and classical AdaBoost algorithm.