An angle-based neighborhood graph classifier with evidential reasoning

Abstract

A classification approach called angle-based neighborhood graph (ANG) is proposed in this paper, which can flexibly define the neighborhood of a given query sample based on the geometrical relation established using an angle parameter. The proposed ANG is geometrically intuitive and can be readily implemented. Compared with the traditional neighborhood graph classifiers, ANG can adjust the size of the neighborhood by tuning the angle parameter to obtain better classification accuracy. To deal with the parameter selection in ANG, an evidential reasoning based approach is proposed. Experimental results are provided for comparing ANG and the traditional neighborhood graph classifiers, including Gabriel Graph (GG), Relative Neighborhood Graph (RNG), β skeletons, and adaptive weighted k nearest neighbors classifiers. It can be concluded that ANG is a simple yet flexible and effective classifier, and the evidential reasoning based parameter selection approach for ANG is also effective.