Efficient landmark-based candidate generation for kNN queries on road networks

The k nearest neighbor (kNN) query on road networks finds the k closest points of interest (POIs) by network distance from a query point. A past study showed that a kNN technique using a simple Euclidean distance heuristic to generate candidate POIs significantly outperforms more complex techniques. While Euclidean distance is an effective lower bound when network distances represent physical distance, its accuracy degrades greatly for metrics such as travel time. Landmarks have been used to compute tighter lower bounds in such cases, however past attempts to use them in kNN querying failed to retrieve candidates efficiently. We present two techniques to address this problem, one using ordered Object Lists for each landmark and another using a combination of landmarks and Network Voronoi Diagrams (NVDs) to only compute lower bounds to a small subset of objects that may be kNNs. Our extensive experimental study shows these techniques (particularly NVDs) significantly improve on the previous best techniques in terms of both heuristic and query time performance.