MB+tree: A Dynamically Updatable Metric Index for Similarity Search

Abstract

One of the common query patterns is to find approximate matches to a given query object in a large database. This kind of query processing is referred as similarity search in a metric space. In this paper, we propose a new metric index MB+tree, called Metric B+tree, which supports near neighbour searching in a generic metric space. MB+tree is aimed at reducing both the number of I/O accesses and the number of distance calculations for similarity search in large databases, while allowing dynamic data updates. In this paper, we show that a B+tree, with an auxiliary tree, can be used as a metric index. Unlike other multidimensional (spatial) access methods, using our approach, we can partition data into disjoint partitions while building/maintaining a metric index, which can lead to a significant cost reduction since the number of metric sub-spaces to be searched is reduced. In order to use MB+tree, a slicing value is proposed. With the slicing value, in addition to space division information, a near neighbour searching can be systematically converted to a range search in B+tree. Several different slicing values are considered namely, one-focus-point scheme and two-focus-point scheme. We also conducted extensive experimental studies using synthetic data. Results are reported in this paper.