Abstract:

Kernel discriminants are greatly appreciated because 1) they may express nonlinear boundaries between classes and 2) they permit to visualize graphically the data points belonging to di®erent classes. One such method is called GDA (Generalized Discriminant Analysis), however it operates on a kernel matrix of size NxN and is for large N prohibitive. We illustrate how this method works in a real situation when dealing with relatively large data. We consider a set of predictors of erosion risk in the Kefallinia island categorized into 5 classes of erosion risk (together N=3422 data items). We argue that a proper preparation of appropriate learning samples can greatly speed up the evaluations and result in good generalization properties. Our concern is to ¯nd appropriate data for learning. This is done by a kind of sieve algorithm.