Clustering-based Nearest Neighbor Searching

Abstract

This paper proposes a Clustering-based Nearest Neighbor Search algorithm (CNNS) for high dimensional data. Different from existing approaches that are based on rigid-grid partition to develop data access structure, CNNS creates indexing structures according to data inherent distribution, with help of a progressive-styled clustering operation. The grids produced in this way adapt to data natural contours. CNNS is characterized with dataset reduction and dimension reduction. And parameterization heuristics are given to bring computation ease to CNNS. Empirical evidence on real datasets demonstrates the fine performance of CNNS.