PL-Tree:An Efficient Access Method for High-Dimensional Data

The quest for processing spatial data in high-dimensional space has resulted in a number of innovative indexing mechanisms. Most of the early methods index data according to their geometric relationships. While they have success to a certain degree, these methods, unfortunately, either lack efficiency, particularly at a higher dimensionality, or are too complex to implement. These drawbacks have made the geometry-based indexing methods less desirable in large-scale applications involving high-dimensional data. To conquer these problems, we introduce a new algebra-based method for indexing high-dimensional data. Our method first partitions the original data space into hypercubes, and then labels each object in a hypercube using a bijective pairing function. All objects in the same hypercube have the same label and vice versa. The bijective pairing function provides the hypothesis to map a high-dimensional vector into a scalar value. This partition and label process continues recursively, until each hypercube contains a pre-determined number of data points. From a structural point of view, our method forms a tree of labels with a number of children for each parent node which we call a PL-tree. Hypercubes in the PL-tree are indexed by labels instead of high-dimensional vectors. In this paper, we present algorithms to construct a PL-tree index and algorithms to carry out the range queries. We have also done experiments that compare the performance of PL-tree with some popular indexing methods.