Deterministic feature selection for regularized least squares classification

Abstract

We introduce a deterministic sampling based feature selection technique for regularized least squares classification. The method is unsupervised and gives worst-case guarantees of the generalization power of the classification function after feature selection with respect to the classification function obtained using all features. We perform experiments on synthetic and real-world datasets, namely a subset of TechTC-300 datasets, to support our theory. Experimental results indicate that the proposed method performs better than the existing feature selection methods. © 2014 Springer-Verlag.