On Two Principles of Designing Multi-Dimensional Access Methods

This paper addresses the problem of supporting region queries over a dynamically growing set of data in feature spaces with many dimensions. It formulates two principles of indexing data in high-dimensional feature spaces and presents an approach to high-dimensional indexing with these principles in mind. The combined goal of the two principles is to increase the density of index regions, which minimizes the number of accesses to persistent storage. In order to impose a desirable behavior on traditional multi-dimensional access methods that normally do not exhibit this behavior in high-dimensional spaces, the proposed approach applies a memory-resident structure with a special kind of space partitioning. The two principles are used not only as design objectives, but also as means of establishing the plausibility of the approach and experimental results. The results provide important evidence suggesting that a greater adherence to these principles leads to an improved retrieval performance.