Fast nearest neighbor search on road networks

Nearest neighbor (NN) queries have been extended from Euclidean spaces to road networks. Existing approaches are either based on Dijkstra-like network expansion or NN/distance precomputation. The former may cause an explosive number of node accesses for sparse datasets because all nodes closer than the NN to the query must be visited. The latter, e.g., the Voronoi Network Nearest Neighbor (V N³) approach, can handle sparse datasets but is inappropriate for medium and dense datasets due to its high precomputation and storage overhead. In this paper, we propose a new approach that indexes the network topology based on a novel network reduction technique. It simplifies the network by replacing the graph topology with a set of interconnected tree-based structures called SPIE's. An nd index is developed for each SPIE and our new (k)NN search algorithms on an SPIE follow a predetermined tree path to avoid costly network expansion. By mathematical analysis and experimental results, our new approach is shown to be efficient and robust for various network topologies and data distributions.