@article{N. Bedo_Oliveira_Traina_Traina Jr._2018, title={Beyond hit-or-miss: a comparative study of synopses for similarity searching}, volume={9}, url={https://sol.sbc.org.br/journals/index.php/jidm/article/view/1635}, DOI={10.5753/jidm.2018.1635}, abstractNote={<p>A DBMS optimizer module takes its decisions by modeling the query costs upon the distribution of the data space. Cost modeling of similarity queries, however, requires the representation of distancesâ€™ rather than data distributions. Therefore, the finding of a suitable representation (or synopsis) for the distance distribution has a major impact in the optimization of similarity searches. In this study, we evaluate the quality of estimates drawn from five synopses of distinct paradigms regarding two common query criteria. Moreover, we embed the synopses into a new parametric cost model, called Stockpile, for the cost estimation of similarity queries on metric trees. The model uses the synopses estimation for calculating the probability of traversing a metric tree node, which defines the expected number of both disk accesses (I/O costs) and distance calculations (CPU costs). We performed an extensive set of experiments on real-world data sources regarding the estimates of each synopsis (and its parametric variations) by using paired ranking tests. In global terms, three synopses have outperformed their competitors regarding selectivity estimation, whereas two of them have also surpassed the others in the prediction of both I/O and CPU costs with respect to Stockpile model predictions. Additionally, results also indicate the choice of the most suitable synopsis may depend on characteristics of the distance distribution.</p>}, number={1}, journal={Journal of Information and Data Management}, author={N. Bedo, Marcos V. and Oliveira, Daniel de and Traina, Agma J. M. and Traina Jr., Caetano}, year={2018}, month={Jun.}, pages={36} }