A0234
Title: Computational strategies for regression subset selection
Authors: Cristian Gatu - University of Iasi (Romania) [presenting]
M Hofmann - University of Oviedo (Spain)
Ana Colubi - University of Giessen (Germany)
Erricos Kontoghiorghes - Cyprus University of Technology and Birkbeck University of London, UK (Cyprus)
Abstract: Computationally efficient algorithms to compute the regression subsets are presented. They are based on regression trees and employ branch-and-bound techniques and heuristics strategies. The main numerical tool that has been employed is the QR factorization and its modification. This yields in a numerically stable and efficient sub-model estimation procedure. An R package ``lmSubsets'' for regression subset selection is presented. The package aims to provide a versatile tool for subset regression. It also embeds a novel algorithm that selects the best variable-subset model according to a pre-determined search criterion. This performs considerably faster than all-subsets variable selection algorithms that rely on the residual sum of squares only. Further computational time improvements based on a parallel version of the branch-and-bound are discussed.
