An Enhanced Classification System Based on Kernel Principal Component Analysis and Data Complexity Measures

Abstract

Principal component analysis is commonly used as a pre-step before employing a classifier to avoid the negative effect of the dimensionality and multicollinearity. The performance of a classifier is severely affected by the deviations from the linearity of the data structure and noisy samples. In this paper, we propose a new classification system that overcomes the drawback of these crucial problems, simultaneously. Our proposal is relying on the kernel principal component analysis with a proper parameter selection approach with data complexity measures. According to the empirical results, F1, T2 and T3 in AUC, T3 in GMEAN and T2 and T3 in MCC performed better than classical and other complexity measures. Comparison of classifiers showed that Radial SVM performs better in AUC, and KNN performs better in GMEAN and MCC using KPCA with complexity measures. As a result, our proposed system produces better results in various classification algorithms with respect to classical approach.