Performance Characterization of Clusterwise Linear Regression Algorithms

Abstract

Clusterwise linear regression (CLR) is a powerful extension of the conventional linear regression framework when the data complexity exceeds the capability of a single linear model. This article presents the first examination of CLR algorithms developed over the past two decades through randomized large-sample testing. Using a unified framework and carefully controlled data characteristics, a comprehensive and systematic assessment of CLR algorithms were performed. The findings of this study provide potential users with a clear understanding of the various benefits and limitations of selecting the appropriate CLR algorithms for their data. Furthermore, this study has disproved past claims which were concluded based on limited samples, and provides insights to better understand the CLR challenges. Finally, this article identifies areas for improvement that could provide crucial performance and reliability improvement of CLR algorithms.