W = SUBSC(A,F)
Each class in the trainingset A is described by linear subspace of dimensionality F (F>=1), or such that at least a fraction F (F<1) of its varianceis retained. This is realised by calling PCAM(AI,F) or for each subset AI of A (objects of class I). For each class a model is built that assumes that the distances of the objects to the class subspaces follow a one-dimensional distribution.
New objects are assigned to the class of the nearest subspace. Classification by D = B*W, in which W is a trained subspace classifier and B is a testset, returns a dataset D with one-dimensional densities for each of the classes in its columns. The result may be improved in case of multi-class problems by a trained combiner, e.g. W = A*(SUBSC*QDC(,,1e-6))
If F is NaN it is optimised by REGOPTC.
This routine will only use the classes with more than two training objects.
E. Oja, The Subspace Methods of Pattern Recognition, Wiley, New York, 1984.