An efficient procedure to select a near-optimal subset of pattern classification features
This dissertation proposes a feature subset selection method that combines several known techniques. Feature subset selection has been identified as one of the most important processes in various applications, especially in pattern classification problems. However, when too many attributes are involved, training a machine to classify patterns into their respective classes is difficult. Hence, selecting good feature subsets is important. Although numerous methods to select feature subsets have been proposed, no universal solution for this problem exists, and the Exhaustive Search method is not practical. Some heuristic techniques such as Sequential Forward Selection and Sequential Backward Elimination are feasible in terms of speed, but suffer from the effects of the local optima problem. The Exhaustive Search technique guarantees finding the optimal subset, but its computational time complexity is exponential. ^ Hence, we propose to minimize the number of features by combining Sequential Forward Selection, Sequential Backward Elimination, and Branch and Bound techniques. This selection produces a subset that the Exhaustive Search technique can be applied to and produce results in a reasonable amount of time. By combining these techniques, the selected feature set is much better than those from heuristic selections, and the computation speed is much faster than the Exhaustive Search technique alone. The proposed combined feature subset selection technique is tested in an off-line signature verification task, using the University of California Irvine (UCI) Repository data sets, by comparing the proposed method with the Exhaustive Search method. ^
Engineering, Industrial|Artificial Intelligence|Computer Science
Kwang Woon Lee,
"An efficient procedure to select a near-optimal subset of pattern classification features"
(January 1, 2003).
ETD Collection for Pace University.