Institute for Systems Research

Specifically, we used state-of-the-art concepts from morphology, namely the attern spectrum' of a shape, to map each shape to a point in n-dimensional space. Following [19, 36], we organized the n-d points in an R-tree. We showed that the L (= max) norm in the n-d space lower-bounds the actual distance. This guarantees no false dismissals for range queries. In addition, we developed a nearest neighbor algorithm that also guarantees no false dismissals.

Finally, we implemented the method, and we tested it against a testbed of realistic tumor shapes, using an established tumor-growth model of Murray Eden [15]. The experiments showed that our method is up to 27 times faster than straightforward sequential scanning.

However, designing feature extraction functions can be hard. It is relatively easier for a domain expert to assess the similarity/distance of two objects. Given only the distance information though, it is not obvious how to map objects into points.

This is exactly the topic of this paper. We describe a fast algorithm to map objects into points in some k- dimensional space ( k is user-defined), such that the dis-similarities are preserved. There are two benefits from this mapping: (a) efficient retrieval, in conjunction with a SAM, ad discussed before and (b) visualization and data-mining: the objects can now be plotted as points in 2-d or 3-d space, revealing potential clusters, correlations among attributes and other regularities that data-mining is looking for.

We introduce an older method from pattern recognition, namely, multi-Dimentional Scaling (MIDS) [Tor52]; although unsuitable for indexing, we use it as yardstick for our method. Then, we propose a much faster algorithm to solve the problem in hand, while in addition it allows for indexing. Experiments on real and synthetic data indeed show that the proposed algorithm is significantly faster than MIDS, (being linear, as opposed to quadratic, on the database size N), while it manages to preserve distances and the overall structure of the data-set.

Finally, we implemented our method on a network of workstations and we compared the experimental and the theoretical results. The conclusions are that (a) our formula for the response time is accurate (the maximum relative error was 30%; the typical error was in the vicinity of 10-15%) (b) the Hilbert-based declustering consistently outperforms a random declustering (c) most importantly, although the optimal chunk size depends on several factors (database size, size of the query, speed of the network), a safe choice for it is 1 page (whichever is the page size of the operating system). We show analytically and experimentally that a chunk size of 1 page gives either optimal or close to optimal results, for a wide range of the parameters.

Armed with this tool, we then show its practical use in predicting the performance of spatial access methods, and specifically of the R-trees. We provide the first analysis of R- trees for skewed distributions of points: We develop a formula that estimates the number of disk accesses for range queries, given only the fractal dimension of the point set, and its count. Experiments on real data sets show that the formula is very accurate: the relative error is usually below 5%, and it rarely exceeds 10%.

We believe that the fractal dimension will help replace the uniformity and independence assumptions, allowing more accurate analysis for any spatial access method, as well as better estimates for query optimization on multi-attribute queries.

For a static database, the proposed ordering achieves superior packing, outperforming older packing methods [25], and the best dynamic method (R*-trees [3]). The savings are as high as 36% on real data.

more importantly, we introduce a dynamic variation, the Hilbert R- tree: : Given the ordering, every node has a well- defined set of sibling nodes; thus, we can deploy the deferred splitting algorithms of the B* -trees. By adjusting the split policy, the Hilbert R-tree can achieve as high utilization as desired. We show that a '3-to-4' split policy achieves good results, consistently outperforming the R* -trees, with up to 28% savings on real data.