Generating seeded trees from data sets

Abstract

In this paper we study the problem of how to perform spatial joins between two data sets with no pre-computed spatial indices. No techniques appear to exist to date that specifically target this problem. Our solution is also useful in the context of query optimization for complex spatial queries. In addition, we demonstrate that simple sampling techniques can be effective in reducing spatial join costs.

We extend the work in [LR94, LR95] and introduce the bootstrap-seeding technique, which allows seeded trees to be constructed directly from input data sets. We can thus dynamically construct two seeded trees for two data sets and perform a spatial join between them. The task of bootstrap-seeding comprises the subtasks of determining the number and the contents of the slots, and constructing the tree. Simple sampling techniques are used to determine the slot contents efficiently.

Our experiments show that spatial joins using our methods are very comparable in performance to that of joins between the same data sets with pre-computed R-trees, and confirm the viability of our method. When joining two data sets with different sizes, our studies suggest that it would be beneficial to bootstrap an initial seeded tree for the smaller data set, and then to construct a seeded tree for the larger data set using copy-seeding and the seed level filtering technique.

This work was supported in part by the Consortium for International Earth Science Information Networking.