Abstract
In this work an R-tree variant, which uses minimum volume covering ellipsoids instead of usual minimum bounding rectangles, is presented. The most significant aspects, which determine R-tree index structure performance, is an amount of dead space coverage and overlaps among the covering regions.
Intuitively, el- lipsoid as a quadratic surface should cover data more tightly, leading to less dead space coverage and less overlaps. Based on studies of many available R-tree variants (especially SR-tree), the eR-tree (ellipsoid R-tree) with ellipsoidal regions is proposed.
The fo- cus is put on the algorithm of ellipsoids construction as it significantly aspects indexing speed and querying performance. At the end, the eR-tree undergoes ex- periments with both synthetic and real datasets.
It proves its superiority especially on clustered sparse datasets.