Parallel Processing Strategies for Skyline Queries Tolerant to Outliers

This paper deals with the problem of efficiently computing the most preferred objects, in the sense of Pareto ordering, in a set of multidimensional objects involving outliers. These outliers, for instance, odd data from ad sales websites, can hide interesting points, so it is important to limit their impact. To do this, we propose a graded Skyline based on the typicality of each object. Such an approach allows to supply an a-cut of the skyline containing both nondominated points, and almost nondominated but more representative points. The computational cost of a naive algorithm performing this task is very high, especially in high-dimensional spaces, due to the large number of comparisons between objects. Since common skyline optimizations cannot be used in our approach, we propose an implementation on parallel architectures, CUDA and OpenMP. Experimental results on both real-world and synthetic data sets show that this processing strategy makes the approach scalable. (C) 2018 Wiley Periodicals, Inc.