MDSC

Trainable classifier for Manhatten Dissimilarity Space

    W = MDSC(A,R,CLASSF)
    W = A*MDSC([],R,CLASSF)
    W = A*MDSC(R,CLASSF)
    W = A*MDSC(CLASSF)
    D = X*W

Input
A	Dateset used for training
R	Dataset used for representation  or a fraction of A to be used for this.  Default: R = A.
CLASSF	Classifier used in dissimilarity space  Default LIBSVC([],[],100)
X	Test set.

Output
W	Resulting, trained feature space classifier
D	Classification matrix

Description

This is a dissimilarity based classifier intended for a feature  respresentation. The training set A is used to compute for every class  its own eigenspace. All eigenvectors are used. A dissimilarity space is  built by the Manhatten (L1, or Minkowsky-1 or city block) distances  between training objects A or test objects X and the representation  objects R after transformation (i.e. rotation) to the eigenspace of  the class of the particular represention object.

Note that Euclidean distances are not affected by rotation, but Manhatten  distances are.

New objects in feature space can be classified by D = X*W or by  D = PRMAP(X,W). Labels can be found by LAB = D*LABELD or LAB = LABELD(D).

See also

datasets, mappings, fdsc,