Kernel Feature Selection

We address the problem of using a kernel spectral criterion function for feature selection. A feature selection paradigm using spectral properties of the affinity matrix of the input data was recently introduced in [11] and which leads to a bilinear interaction between entries of the data sample. Our goal in this paper is to extend the idea of spectral criteria for feature selection to higher order interactions among the data by introducing a kernel operation through which high-order mappings of the original feature vectors can be made possible.