We present an easy-to-use method to do
feature selection that uses Support Vector Machine techniques. The method
consists in a minimization of the number of non-zero components of a linear
model that has a low empirical error on the training set. It can be applied
to two-class problems but also to multi-class, multi-label and regression
ones. We will try to motivate the method and will describe the experiments
which show that it is at leat as good as many existing feature selection
procedures.