Evaluating density-based motion for big data visual analytics

A common strategy for encoding multidimensional data for visual analysis is to use dimensionality reduction techniques that project data with a very large number of objects and dimensions from higher dimensions onto a lower-dimensional space. In visual analytics tasks, the density of the multidimensional clusters can strongly affect how these clusters are perceived. However, this feature can be lost when that dataset is projected into a 2D space, adversely affecting the effectiveness of visual analytics tasks. Thus, it makes sense to preserve, as far as possible, information about the density during the dimensionality reduction. This paper is a study of motion-enhanced cluster perception where the clusters are shown in 2D scatterplots and cluster density is mapped to the motion of the individual constituent points. We consider different types of density-based motion, where the magnitude of the motion is directly related to the density of the clusters. We conducted a series of user studies with large datasets to investigate how motion is a powerful perceptual cue well-suited for grouping or segmenting types during perceptual tasks. We found that the use of motion enabled users to be easily able to distinguish between clusters with different densities. The amount of visual change per unit time was different for the different motions, and we describe the ranges and thresholds for each of them. Specifically, we looked at two projection techniques that output 2D scatterplots for a range of data analysis tasks. We focus on high-dimensional, real-world datasets that might require analyses involving cluster identification, similarity seeking, and cluster ranking tasks. Our results indicate that incorporating density-based motion into visualization analytics systems effectively enables the exploration and analysis of multidimensional datasets.