When Is Nearest Neighbors Indexable?
In this paper, we consider whether traditional index structures are effective in processing unstable nearest neighbors workloads. It is known that under broad conditions, nearest neighbors workloads become unstable–distances between data points become indistinguishable from each other. We complement this earlier result by showing that if the workload for your application is unstable, you are not likely to be able to index it efficiently using (almost all known) multidimensional index structures. For a broad class of data distributions, we prove that these index structures will do no better than a linear scan of the data as dimensionality increases.
Our result has implications for how experiments should be designed on index structures such as R-Trees, X-Trees and SR-Trees: Simply put, experiments trying to establish that these index structures scale with dimensionality should be designed to establish cross-over points, rather than to show that the methods scale to an arbitrary number of dimensions. In other words, experiments should seek to establish the dimensionality of the dataset at which the proposed index structure deteriorates to linear scan, for each data distribution of interest; that linear scan will eventually dominate is a given.
An important problem is to analytically characterize the rate at which index structures degrade with increasing dimensionality, because the dimensionality of a real data set may well be in the range that a particular method can handle. The results in this paper can be regarded as a step towards solving this problem. Although we do not characterize the rate at which a structure degrades, our techniques allow us to reason directly about a broad class of index structures, rather than the geometry of the nearest neighbors problem, in contrast to earlier work.
KeywordsQuery Processing Index Structure Query Point Distance Distribution Indexing Theorem
Unable to display preview. Download preview PDF.
- 1.Beckmann, N., Kriegel, H.-P., Schneider, R., Seeger, B.: The R*-Tree: An Efficient and Robust Access Method for Points and Rectangles. In: Proc. SIGMOD, pp. 322–331 (1992)Google Scholar
- 2.Berchtold, S., Keim, D.A., Kriegel, H.-P.: The x-tree: An Index Structure for High-Dimensional Data. In: Proc. VLDB, pp. 28–39 (1996)Google Scholar
- 3.Beyer, K., Goldstein, J., Ramakrishnan, R., Shaft, U.: When Is Nearest Neighbors Meaningful? In: Proc. ICDT (1999)Google Scholar
- 4.Gionis, A., Indyk, P., Motwani, R.: Similarity search in high dimensions via hashing. In: Proc. VLDB, pp. 518–529 (1999)Google Scholar
- 5.Goldstein, J.: Improved Query Processing and Data Representation Techniques. Ph.D. Thesis, Univ. of Wisconsin-Madison (1999)Google Scholar
- 6.Guttman, A.: R-Trees: A Dynamic Index Structure for Spatial Searching. In: Proc. SIGMOD, pp. 47–57 (1984)Google Scholar
- 7.Hellerstein, J.M., Koutsoupias, E., Papadimitriou, C.H.: On the analysis of indexing schemes. In: Proc. PODS, pp. 249–256 (1997)Google Scholar
- 8.Katayama, N., Satoh, S.: The SR-tree: An Index Structure for High-Dimensional Nearest Neighbor Queries. In: Proc. SIGMOD, pp. 369–380 (1997)Google Scholar
- 9.Lin, K.-I., Jagadish, H.V., Faloutsos, C.: The TV-Tree – An Index Structure for High-Dimensional Data. VLDB J.: Special Issue on Spatial Database Systems 3/4, 517–542 (1994)Google Scholar
- 10.Robinson, J.T.: The K-D-B Tree: A Search Structure for Large Multi-dimensional Dynamic Indexes. In: Proc. SIGMOD, pp. 10–18 (1981)Google Scholar
- 11.Sellis, T.K., Roussopoulos, N., Faloutsos, C.: The R+-Tree: A Dynamic Index for Multi-Dimensional Objects. In: Proc. VLDB, pp. 507–518 (1987)Google Scholar
- 12.Shaft, U.: Database Support for Queries by Image Content. Ph.D. Thesis, Univ. of Wisconsin-Madison (2002)Google Scholar
- 13.Smith, J.R.: Query vector projection access method. In: Storage and Retrieval for Image and Video Databases, vol. VII, pp. 511–522 (1998)Google Scholar
- 14.White, D.A., Jain, R.C.: Similarity Indexing with the SS-tree. In: Proc. ICDE, pp. 516–523 (1996)Google Scholar