NM-Tree: Flexible Approximate Similarity Search in Metric and Non-metric Spaces

Abstract

So far, an efficient similarity search in multimedia databases has been carried out by metric access methods (MAMs), where the utilized similarity measure had to satisfy the metric properties (reflexivity, non-negativity, symmetry, triangle inequality). Recently, the introduction of TriGen algorithm (turning any nonmetric into metric) enabled MAMs to perform also nonmetric similarity search. Moreover, it simultaneously enabled faster approximate search (either metric or nonmetric). However, a simple application of TriGen as the first step before MAMs' indexing assumes a fixed "approximation level", that is, a user-defined tolerance of retrieval precision is preset for the whole index lifetime. In this paper, we push the similarity search forward; we propose the NM-tree (nonmetric tree) --- a modification of M-tree which natively aggregates the TriGen algorithm to support flexible approximatenonmetric or metric search. Specifically, at query time the NM-tree provides a user-defined level of retrieval efficiency/precision trade-off. We show the NM-tree could be used for general (non)metric search, while the desired retrieval precision can be flexibly tuned on-demand.