A nearest neighbor method using bisectors

Abstract

novel algorithm for finding the nearest neighbor was proposed. According to the development of modern technology, the demand is increasing in large-scale datasets with a large number of samples and a large number of features. However, almost all sophisticated algorithms proposed so far are effective only in a small number of features, say, up to 10. This is because in a high-dimensional space many pairs of samples share a same distance. Then the naive algorithm outperforms the others. In this study, we considered to utilize a sequential information of distances obtained by the examined training samples. Indeed, a combinatorial information of examined samples was used as bisectors between possible pairs of them. With this algorithm, a query is processed in O(αβnd) for n samples in a d-dimensional space and for α,β < 1, in expense of a preprocessing time and space in O(n2). We examined the performance of the algorithm. © Springer-Verlag 2004.