Hybrid Search for Optimization of High-Dimensional k-d Trees
Neal Sample, Matthew J. Haines, Mark G. Arnold, Timothy J. Purcell · 1999
While k-d trees have been widely studied and used, their theoretical advantages are often not realized due to ineffective search strategies and generally poor performance in high dimensional spaces. In this paper we outline an effective search algorithm for k-d trees that combines an optimal depth-first branch and bound (DFBB) strategy with a unique method for path ordering and pruning. Our initial method was developed for improving nearest neighbor (NN) search, but has also proven effective for k-NN search and approximate k-NN classification.