Simultaneous classification and feature selection via convex quadratic programming with application to HIV-associated neurocognitive disorder assessment

Support vector machines (SVMs), that utilize a mixture of the L-1-norm and the L-2-norm penalties, are capable of performing simultaneous classification and selection of highly correlated features. These SVMs, typically set up as convex programming problems, are re-formulated here as simple convex quadratic minimization problems over non-negativity constraints, giving rise to a new formulation the pq-SVM method. Solutions to our re-formulation are obtained efficiently by an extremely simple algorithm. Computational results on a range of publicly available datasets indicate that these methods allow greater classification accuracy in addition to selecting groups of highly correlated features. These methods were also compared on a new dataset assessing HIV-associated neurocognitive disorder in a group of 97 HIV-infected individuals. (C) 2010 Elsevier B.V. All rights reserved.