On Kernel Principal Component Regression with Covariance Inflation Criterion for Model Selection

Abstract

This paper introduces Kernel Principal Component Regression (PCR) with the Covariance Inflation Criterion (CIC) for model order selection. The relation to Kernel Ridge Regression (RR) and other 'kernel' regression techniques is given and two benchmark problems demonstrate the comparable performance of CIC to cross-validation techniques. In all reported experiments CIC provides the models with equal performance in comparison to Kernel RR. Moreover, on a significant real world application, Kernel PCR with CIC resulted in smaller model compared to Kernel PCR with the cross-validation technique employed for the selection of principal components.


Go back