Static-to-Dynamic Transformation for Metric Indexing Structures

Abstract

In this paper, we study the well-known algorithm of Bentley and Saxe in the context of similarity search in metric spaces. We apply the algorithm to existing static metric index structures, obtaining dynamic ones. We show that the overhead of the Bentley-Saxe method is quite low, and because it facilitates the dynamic use of any state-of-the-art static index method, we can achieve results comparable to, or even surpassing, existing dynamic methods. Another important contribution of our approach is that it is very simple--an important practical consideration. Rather than dealing with the complexities of dynamic tree structures, for example, the core index can be built statically, with full knowledge of its data set.