A meta-indexing method for fast probably approximately correct nearest neighbor searches

Abstract In this paper we present an indexing method for probably approximately correct nearest neighbor queries in high dimensional spaces capable of improving the performance any index whose degrades with increased dimensionality query space. The basic idea is quite simple: use SVD to concentrate variance inter-element distance a lower space, Ξ. We do space and then “peek” forward from by gathering all elements less than $d_{\Xi }(1+\zeta \sigma _{\Xi }^{2})$ d Ξ ( 1 + ζ σ 2 ) , where d Ξ Ξ, $\sigma }^{2}$ data ζ parameter. All thus collected form tentative set T which scan using complete feature find point closest query. advantages are that (1) it can be built on top virtually (2) build model distribution error precise enough allow designing compromise between speed. show improvement obtain SUN base.