Abstract

Near-neighbor search is an increasingly important operation for queries over multimedia, text, and other non-standard datatypes. In large databases, near-neighbor searches must be enhanced by indexed retrieval for efficiency. In this paper, we present a detailed analysis of three proposals for near-neighbor search: one based on the R-tree, and two which motivated the invention of new trees, namely the SS-tree and SR-tree. We find that while the new trees do improve performance, the reason for their improvement comes mostly from a new Penalty metric, and not from a variety of other details in their implementation. Our analysis was done using a Generalized Search Tree, which both allowed us to easily do a fair comparison, and also provided the framework for a clearer analysis of the issues at hand. 1 Introduction Near-neighbor search is a technique required by a variety of new applications. Given a particular data item, near-neighbor search finds the k closest matches to the item and ou...

Citations