Selectivity Estimation of High Dimensional Window Queries Via Clustering

Query optimization is an important functionality of modern database systems and often based on estimating the selectivity of queries before actually executing them. Well-known techniques for estimating the result set size of a query are sampling and histogram-based solutions. Sampling-based approaches heavily depend on the size of the drawn sample which causes a trade-off between the quality of the estimation and the time in which the estimation can be executed for large data sets. Histogram-based techniques eliminate this problem but are limited to low-dimensional data sets. They either assume that all attributes are independent which is rarely true for real-world data or else get very inefficient for high-dimensional data. In this paper we present the first multivariate parametric method for estimating the selectivity of window queries for large and high-dimensional data sets. We use clustering to compress the data by generating a precise model of the data using multivariate Gaussian distributions. Additionally, we show efficient techniques to evaluate a window query against the Gaussian distributions we generated. Our experimental evaluation shows that this approach is significantly more efficient for multidimensional data than all previous approaches