Matching pursuit kernel fisher discriminant analysis

We derive a novel sparse version of Kernel Fisher Discriminant Analysis (KFDA) using an approach based on Matching Pursuit (MP). We call this algorithm Matching Pursuit Kernel Fisher Discriminant Analysis (MPKFDA). We provide generalisation error bounds analogous to those constructed for the Robust Minimax algorithm together with a sample compression bounding technique. We present experimental results on real world datasets, which show that MPKFDA is competitive with the KFDA and the SVM on UCI datasets, and additional experiments that show that the MPKFDA on average outperforms KFDA and SVM in extremely high dimensional settings.