A Dynamic Pivoting Algorithm Based on Spatial Approximation Indexes

Metric indexes aim at reducing the amount of distance evaluations carried out when searching a metric space. Spatial approximation trees (sa-trees for short), in particular, are efficient data structures, which have shown to be competitive in metric spaces of medium to high difficulty, or queries with low selectivity. Sa-trees can be also made dynamic, and can use the available space to improve the query performance adding pivot information. In this paper we show how the pivot information added to dynamic sa-trees can be used to a full extent, in order to improve the search performance. The result is a technique that allows one to traverse a dynamic sa-tree without necessarily comparing all traversed nodes against the query object. As a result, the novel algorithm makes a much better use of the available space, yielding a saving of distance computations of about 10% to 70%, compared with previous sa-tree schemes that use pivot information.