Variable Selection for Kernel Classifiers: A Feature-to-Input Space Approach

An aspect of kernel classifiers which complicates variable selection is the implicit use of the transformation function Φ. This function maps the space in which the data cases reside, the so-called input space (\(\mathcal{X}\)), to a higher dimensional feature space (\(\mathcal{F}\)). Variable selection in \(\mathcal{F}\) is a difficult problem, while variable selection in \(\mathcal{X}\) is mostly inadequate. We propose an intermediate kernel variable selection approach which is implemented in \(\mathcal{X}\) while also accounting for the fact that kernel classifiers operate in \(\mathcal{F}\).