Golub-Kahan Regression for Variable Selection in Statistical Modeling

In linear models usually the solutions are required to be sparse since they identify the most important factors that influence the response of the experiment. In this work we apply a Krylov-based method in statistical modeling. More precisely, we describe the direct method skinny Golub-Kahan regression which behaves well for the under study statistical models in terms of the required sparsity in the computed solutions. This method is based only on matrix vector products operations and inner products of vectors as well. Therefore it is computationally effective and according to the presented experiments it is highly recommended for datasets possessing tall thin data matrix.