Abstract
The M-tree and its variants have been proved to provide an efficient similarity search in database environments. In order to further improve their performance, in this paper we propose an extension of the M-tree family, which makes use of nearest-neighbor (NN) graphs.
Each tree node maintains its own NN-graph, a structure that stores for each node entry a reference (and distance) to its nearest neighbor, considering just entries of the node. The NN-graph can be used to improve filtering of non-relevant subtrees when searching (or inserting new data).
The filtering is based on using ?sacrifices? ? selected entries in the node serving as pivots to all entries being their reverse nearest neighbors (RNNs). We propose several heuristics for sacrifice selection; modified insertion; range and kNN query algorithms.
The experiments have shown the M-tree (and variants) enhanced by NN-graphs can perform significantly faster, while keeping the construction cheap.