Acceleratingk-nearest-neighbor searches

Abstract

The search for whichkpoints are closest to a given probe point in a space ofNknown points, the `k-nearest-neighbor' or `KNN' problem, is a computationally challenging problem of importance in many disciplines, such as the design of numerical databases, analysis of multi-dimensional experimental data sets, multi-particle simulations and data mining. A standard approach is to preprocess the data into a tree and make use of the triangle inequality to prune the search time to the order of the logarithm ofNfor a single nearest point in a well balanced tree. All known approaches suffer from the `curse of dimensionality', which causes the search to explore many more branches of the tree than one might wish as the dimensionality of the problem increases, driving search times closer to the order ofN. Looking forknearest points can sometimes be done in approximately the time needed to search for one nearest point, but more often it requiresksearches because the results are distributed widely. The result is very long search times, especially when the search radius is large andkis large, and individual distance calculations are very expensive, because the same probe-to-data-point distance calculations need to be executed repeatedly as the top of the tree is re-explored. Combining two acceleration techniques was found to improve the search time dramatically: (i) organizing the search into nested searches in non-overlapping annuli of increasing radii, using an estimation of the Hausdorff dimension applicable to this data instance from the results of earlier annuli to help set the radius of the next annulus; and (ii) caching all distance calculations involving the probe point to reduce the cost of repeated use of the same distances. The result of this acceleration in a search of the combined macromolecular and small-molecule data in a combined six-dimensional database of nearly 900 000 entries has been an improvement in the overall time of the searches by one to two orders of magnitude.