Avoiding the interpolation inaccuracy in nearest feature line classifier by spectral feature analysis

Abstract

In nearest feature line approach, the representational capacity of a given training set is generalized by defining feature lines passing through each pair of samples belonging to the same class. This technique is shown to provide superior performance on various classification problems than the nearest neighbor approach. From the performance point of view, the major weakness of this technique is the interpolation inaccuracy which occurs when a feature line passes through samples that are far away from each other. Several variants are recently proposed to avoid this weakness. In this study, we follow a different path and propose to transform the training data of different classes into separate clusters before applying nearest feature line classifier. Spectral clustering based transformation is used for this purpose and it is shown that the accuracies achieved by both the nearest feature line and the shortest feature line segment approach which is the most recent variant of the nearest feature line technique are improved. (c) 2013 Elsevier B.V. All rights reserved.