Pseudo Nearest Centroid Neighbor Classification

Abstract

In this paper, we propose a new reliable classification approach, called the pseudo nearest centroid neighbor rule, which is based on the pseudo nearest neighbor rule (PNN) and nearest centroid neighborhood (NCN). In the proposed PNCN, the nearest centroid neighbors rather than nearest neighbors per class are first searched by means of NCN. Then, we calculate k categorical local mean vectors corresponding to k nearest centroid neighbors, and assign the weight to each local mean vector. Using the weighted k local mean vectors for each class, PNCN designs the corresponding pseudo nearest centroid neighbor and decides the class label of the query pattern according to the closest pseudo nearest centroid neighbor among all classes. The classification performance of the proposed PNCN is evaluated on real data sets in terms of the classification accuracy. The experimental results demonstrate the effectiveness of PNCN over the competing methods in many practical classification problems.