MAMCost: Global and Local Estimates leading to Robust Cost Estimation of Similarity Queries

Abstract

This paper presents an effective cost model to estimate the number of disk accesses (I/O cost) and the number of distance calculations (CPU cost) to process similarity queries over data indexed by metric access methods. Two types of similarity queries were taken into consideration: range and k-nearest neighbor queries. The main point of the cost model is considering not only global parameters of the data set but also the local data distribution. The model takes advantage of the intrinsic dimension of the data set, estimated by its correlation fractal dimension. Experiments were performed on real and synthetic data sets, with different sizes and dimensions, in order to validate the proposed model. They confirmed that the estimations are accurate, within the range achieved by real queries.