Learning kernel parameters for kernel Fisher discriminant analysis

Abstract

Kernel Fisher discriminant analysis (KFDA) faces the problem of kernel parameters selection. A novel KFDA kernel parameters optimization criterion is presented for maximizing the uniformity of class-pair separabilities and class separability in kernel space simultaneously. The presented criterion is also applied to the kernel parameters selection of spectral regression kernel discriminant analysis (SRKDA). Minimum distance classifier, k nearest neighbor (kNN) classifier, and naive Bayes classifier are used to evaluate the feature extraction performance. Experiments on fourteen benchmark multiclass data sets show that, comparing with the criterion for merely maximizing the class separability in kernel space, the presented criterion can search the optimum KFDA kernel parameters more accurately, and do better in SRKDA kernel parameters selection.