Abstract. Supervised kernel-Principal Component Analysis (S-kPCA) is a me thod for producing discriminative feature spaces that provide nonlinear decision regions, well-suited for handling real-world problems. The presented paper proposes a modification to the original S-kPCA concept, which is aimed at improving class-separation in resulting feature spaces. This is accomplished by identifying outliers (understood here as misclassified samples) and by an appropriate reformulation of the original S-kPCA problem. The proposed idea is to replace binary class labels that are used in the original method, by real-valued ones, derived using sample-relabeling scheme aimed at preventing potential data classification problems. The postulated concept has been tested on three standard pattern recognition datasets. It has been shown that classification performance in feature spaces derived using the introduced methodology improves by 4–16% with respect to the original S-kPCA method, depending on a dataset.
