An empirical evaluation of four data visualization techniques for displaying short news text similarities

An experiment was conducted comparing user performance on four data visualization techniques—an unstructured display condition consisting of a random one-dimensional (1D) list and three proximity-based representations including a 1D list ranked by a greedy nearest-neighbor algorithm and two 2D spatial visualizations using the ISOMAP layout algorithm and multidimensional scaling (MDS). Eighty-one participants completed an information retrieval task where the visualization techniques were used to display a corpus consisting of 50 short news texts. Human pairwise similarity judgments for this corpus were used to create the three proximity-based displays. Results demonstrated an advantage in accuracy, the number of documents accessed, and, to a lesser extent, subjective confidence in these displays over the Random List condition and in the 2D over the 1D displays. Similar, but smaller, advantages were observed in the MDS display over ISOMAP however none of these pairwise comparisons were statistically significant. A sequential analysis of participant actions in terms of the proximity of document representations accessed provided some explanation for variations in performance between the displays as well as indicating strategic differences in interactions particularly between visualizations of different dimensionality.