Abstract:

This paper addresses the problem of variable ranking for sup- port vector regression. We use a novel wrapper algorithm for feature selection, using bounds of support vector machines in re- gression with kernel functions.The ranking criteria that we pro- posed are based on leave-one-out bounds such us radius margin and span estimate bound, we have applied these criteria to a classical search-space algorithm such us backward elimination. All these algorithms have been compared on toy problems and real-world air pollution data sets. The results obtained by the SVR were compared with those obtained by SVR after selection. The comparison of these methods is in favor of regression after variable selection considering the radius margin as a criterion.