A class-based evaluation approach to assess multidimensional projections

Abstract

Multidimensional projection techniques have been widely used to visually explore datasets due to their ability to generate representations that preserve similarity relations of data points into lower dimensional spaces. To evaluate if the embedded space reflects high-dimensional structures, measures are usually employed to return a quality score of the whole projection. In contrast to this idea, we evaluate the embedded layouts by assessing each class of the datasets at a time by using well-known quality measures. In addition, we propose assessing multidimensional projection techniques using ROC curves. Experimental results on two datasets show that our approach can be useful to discover how classes interact each other by using different visualization techniques and how close-related they are without thoroughly exploring the layouts. ROC curves proved to be a good measure for analyzing projection techniques and can give highly valuable feedback to users when exploring multidimensional data.