A Pipeline for Tailored Sampling for Progressive Visual Analytics

Abstract

Progressive Visual Analytics enables analysts to interactively work with partial results from long-running computations early on instead of forcing them to wait. For very large datasets, the first step is to divide that input data into smaller chunks using sampling, which are then passed down the progressive analysis pipeline all the way to their progressive visualization in the end. The quality of the partial results produced by the progression heavily depends on the quality of these chunks, that is, chunks need to be representative of the dataset. Whether or not a sampling approach produces representative chunks does however depend on the particular analysis scenario. This stands in contrast to the common use of random sampling as a ''one-size-fits-most'' approach in PVA. In this paper, we propose a sampling pipeline and its open source implementation which can be used to tailor the used sampling method for an analysis scenario at hand. This pipeline consists of three configurable steps - linearization, subdivision, and selection - and for each, we propose exemplar operators. We then demonstrate its utility by providing tailored samplings for three distinct scenarios.