A Method of Kernel Fisher Discriminant for Multi-class Classification

Kernel Fisher discriminant analysis (KFD) has good performance in practice as a classification method. However, KFD is initially developed for binary classification. To solving multi-class classification problems, multi-class KFD (MKFD) was designed to minimize total deviation. By Lagrange multiplier method, MKFD was transformed to be a quadratic optimization problem that can avoid solving eigenproblem and be less numerical demanding relatively. Moreover it is shown that MKFD is a direct generalization of the binary classification. Finally the performance of MKFD was tested on the benchmark datasets in experiments. The results support usefulness of MKFD, compared with other methods such as support vector machines