Kernel Parameter Optimization for KFDA Based on the Maximum Margin Criterion

Kernel parameters optimization is one of the most challenging problems on kernel Fisher discriminant analysis (KFDA). In this paper, a simple and effective KFDA kernel parameters optimization criterion is proposed on the basis of the maximum margin criterion (MMC) that maximize the distances between any two classes. Actually, this MMC-based criterion is applied to the kernel parameters optimization on KFDA and KFDA with Locally Linear Embedding affinity matrix (KFDA-LLE). It is demonstrated by the experiments on six real-world multiclass datasets that, in comparison with two other criteria, our MMC-based criterion can detect the optimal KFDA kernel parameters more accurately in the cases of both RBF kernel and polynomial kernel.