Combining elimination rules in tree-based nearest neighbor search algorithms

Abstract

A common activity in many pattern recognition tasks, image processing or clustering techniques involves searching a labeled data set looking for the nearest point to a given unlabelled sample. To reduce the computational overhead when the naive exhaustive search is applied, some fast nearest neighbor search (NNS) algorithms have appeared in the last years. Depending on the structure used to store the training set (usually a tree), different strategies to speed up the search have been defined. In this paper, a new algorithm based on the combination of different pruning rules is proposed. An experimental evaluation and comparison of its behavior with respect to other techniques has been performed, using both real and artificial data. © 2010 Springer-Verlag Berlin Heidelberg.