L0 zero-norm minimization (Weston,Elisseeff dual method) object
A=L0(H) returns a l0 object initialized with hyperparameters H.
Minimizes the zero norm of the weight vector subject to
separability constraints. Although it is possible to deal with the
non-seperable case (see paper), this is not currently implemented.
Hyperparameters, and their defaults
feat=0 -- number of features to be selected, if feat=0,
then the minimum number of feature will be computed.
output_rank=1 -- whether a ranking is desired, if set to 0 then a
classification is perfomed after feature selection.
find_min=1 -- stop with the minimum number of features if set to 1
otherwise stop as soon as reach feat nonzero features
child=svm -- Set the classifiers to be used at each step
a.rank= -- ranking of features
a.child=svm -- base classifier trained at end of process
d=gen(toy); a=l0; a.feat=20; a.output_rank=1;[r,a]=train(a,d);
a.rank - lists the chosen features in order of importance, using 20 features
d=gen(toy); a=l0; a.feat=0; a.output_rank=1;[r,a]=train(a,d);
a.rank - lists the chosen features in order of importance, using minimum number of features
Reference : Use of the zeronorm with linear models and kernel methods |
Author : J. Weston, A. Elisseeff, B. Schölkopf, and M. Tipping |