FEATSELB

Trainable mapping for backward feature selection

    [W,R] = FEATSELB(A,CRIT,K,T)
    [W,R] = A*FEATSELB([],CRIT,K,T)
    [W,R] = A*FEATSELB(CRIT,K,T)
    [W,R] = FEATSELB(A,CRIT,K,N)
    [W,R] = A*FEATSELB([],CRIT,K,N)
    [W,R] = A*FEATSELB(CRIT,K,N)

Input
A	Dataset
CRIT	String name of the criterion or untrained mapping  (optional; default: 'NN', i.e. 1-Nearest Neighbor error)
K	Number of features to select  (optional; default: return optimally ordered set of all features)
T	Tuning set (optional)
N	Number of cross-validation folds

Output
W	Output feature selection mapping
R	Matrix with step-by-step results of the selection

Description

Backward selection of K features using the dataset A. CRIT sets the  criterion used by the feature evaluation routine FEATEVAL. If the  dataset T is given, it is used as test set for FEATEVAL. Alternatvely a  a number of cross-validation N may be supplied. For K = 0, the optimal  feature set (corresponding to the maximum value of FEATEVAL) is returned.  The result W can be used for selecting features by B*W. In this case,  features are ranked optimally.  The selected features are stored in W.DATA and can be found by +W.  In R, the search is reported step by step as

     R(:,1) : number of features
     R(:,2) : criterion value
     R(:,3) : added / deleted feature

See also

mappings, datasets, feateval, featsellr, featsel, featselo, featself, featseli, featselp, featselm,