Robust principal curves based on maximum correntropy criterion

Chun-Guo Li, Bao-Gang Hu

Abstract: Principal curves are curves which pass through the `middle′ of a data cloud. They are sensitive to variances of data clouds. In this paper, we propose a robust principal curve model - Correntropy based Principal Curve (CPC) model, based on maximum correntropy criterion (MCC). The CPC model approximate the principal curve with k-segments of polygonal line. Employing the half-quadratic technique, the CPC model is optimized in an iteratively way. The CPC model is insensitive to variances and outliers of data clouds. Extensive experiments on synthetic and real-life datasets illustrate the robustness of the CPC model in learning principal curves.