Editing the nearest feature line classifier

Abstract

Two major drawbacks in nearest feature line classifier are the extrapolation and interpolation inaccuracies. The former can easily be counteracted by considering segments instead of lines. However, the solution of the latter problem is more challenging. Recently developed techniques tackle with this drawback by selecting a subset of the feature line segments either during training or testing. In this study, a novel scheme that is based on editing the feature line segments is proposed. It involves two major strategies, namely generalization-based elimination and representation-based elimination. In generalization-based elimination, the benefit and cost of discarding each feature line segment are compared and the segments that contribute to the classification error are discarded. For the implementation of representation-based elimination, a measure of intersection is defined and the longer of each pair of intersecting segments are discarded. Moreover, the distances of the feature line segments to the samples of other classes are also investigated to discard inconsistent segments. The proposed steps are evaluated by applying them in different orders. Experimental results have shown that the proposed approach provides better accuracies when compared to two recently developed extensions of the nearest feature line approach, namely shortest feature line segment and rectified nearest feature line segment.