FEATEVAL

Evaluation of feature set for classification

    J = FEATEVAL(A,CRIT,T)
    J = A*FEATEVAL([],CRIT,T)
    J = A*FEATEVAL(CRIT,T)
    J = FEATEVAL(A,CRIT,N)
    J = A*FEATEVAL([],CRIT,N)
    J = A*FEATEVAL(CRIT,N)

Input
A	input dataset
CRIT	string name of a method or untrained mapping, default 'NN'
T	validation dataset (optional)
N	number of cross-validation folds (optional)

Output
J	scalar criterion value

Description

Evaluation of features by the criterion CRIT, using objects in the  dataset A. The larger J, the better. Resulting J-values are  incomparable over the various methods.  The following methods are supported

     crit='in-in' : inter-intra distance.
     crit='maha-s': sum of estimated Mahalanobis distances.
     crit='maha-m': minimum of estimated Mahalanobis distances.
     crit='eucl-s': sum of squared Euclidean distances.
     crit='eucl-m': minimum of squared Euclidean distances.
     crit='NN'    : 1-Nearest Neighbour leave-one-out
                    classification performance (default).
                    (performance = 1 - error). 
     crit='mad'   : mean absolute deviation (only for regression!)
     crit='mse'   : mean squared error (only for regression!)

For classification problems, CRIT can also be any untrained  classifier, e.g. LDC([],1e-6,1e-6). Then the classification error is  used for a performance estimate. If supplied, the dataset T is used  for obtaining an unbiased estimate of the performance of classifiers  trained with the dataset A. If a number of cross-validations N is  supplied, the routine is run for N times with different training and  test sets generated from A by cross-validation. Results are averaged.  If T nor N are given, the apparent performance on A is used.

See also

datasets, featselo, featselb, featself, featselp, featselm, featrank,