LKC

Trainable linear kernel classifier

    W = LKC(A,KERNEL)
    W = A*LKC([],KERNEL)
    W = A*LKC(KERNEL)

Input
A	Dataset
KERNEL	Mapping to compute kernel by A*MAP(A,KERNEL) or string to compute kernel by FEVAL(KERNEL,A,A) or cell array with strings and parameters to compute kernel by
FEVAL(KERNEL{1},A,A,KERNEL{2:END})	Default: linear kernel (PROXM('P',1))

Output
W	Mapping: Support Vector Classifier

Description

This is a fall-back routine for other kernel procedures like SVC, RBSVC and LIBSVC. If they fail due to optimisation problems they may fall back  to this routine which computes a linear classifier in kernelspace using  the pseudo-inverse of the kernel.

The kernel may be supplied in KERNEL by

• an untrained mapping, e.g. a call to PROXM like W = LIBSVC(A,PROXM('R',1))
• a string with the name of the routine to compute the kernel from A
• a cell-array with this name and additional parameters.  This will be used for the evaluation of a dataset B by B*W or PRMAP(B,W) as  well.

If KERNEL = 0 it is assumed that A is already the kernel matrix (square).  In this also a kernel matrix should be supplied at evaluation by B*W or  PRMAP(B,W).

LKC is basically a two-class classifier. Multi-class problems are solved  in a one-against-rest fashion by MCLASSC. The resulting base-classifiers  are combined by the maximum confidence rule. A better, non-linear  combiner might be QDC, e.g. W = A*(LKC*QDC([],[],1e-6))

See also

mappings, datasets, svc, proxm,