Stress Maps: Analysing Local Phenomena in Dimensionality Reduction Based Visualizations

Challenges in Visual Analytics frequently involve massive repositories, which do not only contain a large number of information artefacts, but also a high number of relevant dimensions per artefact. Dimensionality reduction algorithms are commonly used to transform high-dimensional data into low- dimensional representations which are suitable for visualisation purposes. For example, Information Landscapes visualise high-dimensional data in two dimensions using distance-preserving projection methods. The inaccuracies introduced by such methods are usually expressed through a global stress measure which does not provide insight into localised phenomena. In this paper, we propose the use of Stress Maps, a combination of heat maps and information landscapes, to support algorithm development and optimization based on local stress measures. We report on an application of Stress Maps to a scalable text projection algorithm and describe two categories of problems related to localised stress phenomena which we have identified using the proposed method.