Progressive Parameter Space Visualization for Task-Driven SAX Configuration

Abstract

As time series datasets are growing in size, data reduction approaches like PAA and SAX are used to keep them storable and analyzable. Yet, finding the right trade-off between data reduction and remaining utility of the data is a challenging problem. So far, it is either done in a user-driven way and offloaded to the analyst, or it is determined in a purely data-driven, automated way. None of these approaches take the analytic task to be performed on the reduced data into account. Hence, we propose a task-driven parametrization of PAA and SAX through a parameter space visualization that shows the difference of progressively running a given analytic computation on the original and on the reduced data for a representative set of data samples. We illustrate our approach in the context of climate analysis on weather data and exoplanet detection on light curve data.