A k-means-based and no-super-parametric Improvement of AdaBoost and its Application to Transaction Fraud Detection

Abstract

AdaBoost is a well-known effective boosting algorithm for classification and has achieved successful applications in many fields. The existing studies show that it is very sensitive to noisy points, resulting in a decline of classification performance. We have proposed an improved algorithm called CAdaBoost in order to overcome the weakness. However, our CAdaBoost uses a set of super-parameters. In this paper, we propose a no-super-parametric improvement to CAdaBoost and it is applied to the problem of detecting credit card fraud. Although the performance of this CAdaBoost without super-parameters is a little worse than the original CAdaBoost, it still outperforms others including the original AdaBoost and several existing improvements of AdaBoost. Our design without super-parameters provides a helpful idea for other similar problems.