| |
Feature scaling/selection via SVMs and r^2/w^2 bound.
|
A=R2W2_SEL(H) returns an r2w2_sel object with hyperparameters H.
Hyperparameters, and their defaults
feat=[] -- number of features, (default means
use scaling factors instead)
output_rank=0 -- output rank (if 0, output predicted labels)
slacks=0; -- if slacks=1,optimize ridge as well,
if this is a vector it specifies
which examples share slacks (e.g 3 examples,
slacks=[1 1 2] means first two examples
have a shared slack variable)
scales=1; -- if scales=1, find scaling factors, if 0
do not, if scale is a vector, specifies
sharing of scaling factors as in slacks parameter
max_iter=100 -- maximum number of gradient steps
steep_ascent=0 -- use steepest grad ascent (can be faster when
there are many features)
use_var2=1 -- use var^2w^2 estimate instead of r^2w^2
ker='weighted_linear'-- type of kernel, other choices='weighted_poly'
or 'weighted_rbf'
kerparam -- associated param, e.g degree of poly
optimizer='default' -- choices={default,andre,quadprog,svmlight}
Model
rank -- ranking of the features
child -- decision rule stored in child member (svm)
sigma -- value of all scaling and slacking factors
Example:
d=gen(toy); a=r2w2_sel; 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
| |
Reference : Feature selection for SVMs |
| |
Author : J. Weston, S. Mukherjee, O. Chapelle, M. Pontil, T. Poggio, and V. Vapnik |