A fast algorithm for finding k-nearest neighbors with non-metric dissimilarity

Abstract

Fast nearest neighbor (NN) finding has been extensively studied. While some fast NN algorithms using metrics rely on the essential properties of metric spaces, the others using non-metric measures fail for large-size templates. However in some applications with very large size templates, the best performance is achieved by NN methods based on the dissimilarity measures resulting in a special space where computations cannot be pruned by the algorithms based-on the triangular inequality. For such NN methods, the existing fast algorithms except condensing algorithms are not applicable. In this paper, a fast hierarchical search algorithm is proposed to find k-NNs using a non-metric measure in a binary feature space. Experiments with handwritten digit recognition show that the new algorithm reduces on average dissimilarity computations by more than 90% while losing the accuracy by less than 0.1%, with a 10% increase in memory.