Fast approximate search algorithm for nearest neighbor queries in high dimensions

Abstract

We present a fast approximate nearest neighbor (NN) search index structure called the AB-tree, which uses heuristics to decide whether or not to access a node in the index tree based on the intersecting angle and the weight of the node. The goal of the NN search algorithm presented is to decrease unnecessary node accesses in the search due to overlap among bounding regions in existing index structures. We have observed the following three properties of bounding hyperspheres that motivated the creation of the proposed heuristics for NN search: distance property; radius property; and angle property.