Using Copula and Quantiles Evolution in Prediction of Multidimensional Distributions for Better Query Selectivity Estimation

Abstract

In query optimization theory a selectivity parameter is used by cost query optimizer for early estimating the size of data that satisfies a query condition. It requires some representation of distribution of attribute values. There are many approximate representations of m–d distribution where the copula-based is new one. This approach gives a possibility to take into account the fact of a varying m–d distribution by predicting both a copula and 1–d marginal distributions. In this paper we propose the method of forecasting trajectories of either copula parameters and marginals’ quantiles using time series prediction models. This method is mainly designated for predicting outdated distribution representation what may improve accuracy of selectivity estimation based on such representation. It also may be used for predicting a varying query workload to forecast important regions of data domain. Having detected such regions we may improve there the resolution of distribution representation.