Indexing of location-dependent data in mobile computing environments

Abstract

This thesis conducts research on advanced indexing techniques to improve the performance of location-dependent information services (LDISs). For location-dependent data (LDD), the validity of a data value is dependent on the location of the client retrieving the data value. The retrieval of LDD requires the identification of the region that the client is residing in, given a partition of the geographical space. For example, finding the name of the city that the client is visiting requires identifying which city the client is located in given the city boundaries of a country. D-tree is proposed to efficiently determine the target region by indexing the division of adjacent regions. Compared to the existing approximation methods, such as Minimal Bounding Rectangles (MBRs) in R-tree, D-tree does not index any overlapping region and therefore avoids back-tracking. Compared to decomposition schemes, such as the trapezoidal map, D-tree represents the divisions in the search space and does not introduce any auxiliary line segments to the space. Therefore, the storage cost of D-tree is very low. Simulation results of both synthetical and real datasets show that D-tree outperforms existing indexes significantly in terms of search time. Next, we focus on nearest-neighbor search, which can be answered based on the locations of the data objects in the search space. We propose the grid-partition index which first partitions the whole search space into disjoint sub-spaces, called grid cells, to efficiently narrow down the required search space for a given query point and then for each grid cell index only those objects that are possibly the nearest neighbors to some query point within that grid cell. Since the indexing of objects is more efficient than that of regions and the number of objects indexed by each grid cell is much smaller than the whole population, the search performance is improved. Three partition algorithms have been proposed to partition the original space into disjoint grid cells with the help of a newly defined performance metric, index efficiency. Based on the proposed D-tree and grid-partition index, different kinds of location-dependent queries (LDQs) can be efficiently answered. However, a specific D-tree only serves one kind of LDQ and a grid-partition index is only workable for a nearest-neighbor search. Therefore, a universal index structure is still desirable to answer different LDQs. Motivated by this, a Hilbert-curve index is proposed. To cater for the linear browsing limitation of wireless broadcast systems, the objects are mapped onto a linear space using a space-filling curve, namely the Hilbert curve. Different search algorithms are devised to answer window queries, k-nearest-neighbor queries, and continuous-nearest-neighbor queries. A series of simulation experiments are conducted to evaluate the proposed search algorithms. (Abstract shortened by UMI.)