Um novo método de poda iterativo de máquinas de vetores-suporte de mínimos quadrados

Abstract

Least-Square Support Vector Machines (LSSVM) are solved from a linear system, unlike support vector machines (SVM), which require the application of quadratic programming. So it is preferable in some problems to use LSSVM because the complexity of the resolution will be lower. However, unlike SVM, LSSVM generates non-sparse solutions. A significant disadvantage, given that all training patterns will be support vectors of LSSVM. To overcome such non-sparse solutions, we propose an iterative solution with pruning without causing a loss of efficiency, named as FSLM-LSSVM. The proposed method is compared with two others in the literature, P-LSSVM and IP-LSSVM. Finally, the evaluation of the proposed solution consisted of comparing the averages of the hit rates.