An Improved Wrapper Method for Feature Selection

The wrapper feature selection methods can achieve high classification accuracy, however, its cross-validation scheme in evaluation phase is very expensive in terms of computing resource consumption. In this paper, we propose a new statistical LW-measure which can replace the cross-validation scheme to evaluate feature sets. Furthermore, two improved wrapper algorithms, i.e. sequential forward selection-LW (SFS-LW) and sequential backward selection-LW (SBS-LW), are presented for feature selection, on the basis of combination of LW-measure and sequence search algorithms. Three groups of experiments conducted on two University of California, Irvine (UCI) datasets show that the proposed algorithms can not only obtain the similar classification accuracy to that of the traditional wrapper methods, but also are nearly ten times faster than the traditional ones.