Automatic radial basis STATSSVM optimising the Parzen kernel
[W,KERNEL] = PKSTATSSVC(A,ALF)
This routine provides a radial basis support vector classifier based on STATSSVC. It uses a radial basis kernel supplied by PROXM with a kernel width SIGMA found by PARZENC. The kernel width used is ALF*3*SQRT(2)*SIGMA. This is much faster than the gridsearch used by RBSTATSSVC and performs about equally well.
Note that STATSSVC is basically a two-class classifier and solves multi-class problems by MCLASSC, generating a set of one-against-rest classifiers. PKSTATSSVC, however, supplies a single kernel. For some problems a proper prescaling of the data, e.g. by SCALEM, amy thereby be appropriate.