A centroid k-nearest neighbor method

Abstract

k-nearest neighbor method (kNN) is a very useful and easy-implementing method for real applications. The query point is estimated by its k nearest neighbors. However, this kind of prediction simply uses the label information of its neighbors without considering their space distributions. This paper proposes a novel kNN method in which the centroids instead of the neighbors themselves are employed. The centroids can reflect not only the label information but also the distribution information of its neighbors. In order to evaluate the proposed method, Euclidean distance and Mahalanobis distance is used in our experiments. Moreover, traditional kNN is also implemented to provide a comparison with the proposed method. The empirical results suggest that the propose method is more robust and effective. © 2010 Springer-Verlag.