Efficient Similarity Search in Metric Spaces with Cluster Reduction

Clustering-based methods for searching in metric spaces partition the space into a set of disjoint clusters. When solving a query, some clusters are discarded without comparing them with the query object, and clusters that can not be discarded are searched exhaustively. In this paper we propose a new strategy and algorithms for clustering-based methods that avoid the exhaustive search within clusters that can not be discarded, at the cost of some extra information in the index. This new strategy is based on progressively reducing the cluster until it can be discarded from the result. We refer to this approach as cluster reduction. We present the algorithms for range and kNN search. The results obtained in an experimental evaluation with synthetic and real collections show that the search cost can be reduced by a 13% - 25% approximately with respect to existing methods.