Visualization with Uncertainty

University Of New Mexico, Albuquerque NM

Investigators

Abstract

This research develops visualization methods that aid decision-making by exposing the important data characteristics and their inherent uncertainties in meaningful ways. The goal of visualization is to gain insight and understanding by creating visual representations of complex information. Most visualization methods use some form of classification to eliminate unimportant regions and illuminate interesting ones. The process of classification is inherently uncertain; in general the source data contains noise, data transformations can further introduce and magnify uncertainty. More advanced classification methods rely on approximation models or statistical methods to determine features of interest. While these classification methods can model uncertainty or partial memberships, they typically only provide discrete memberships. Visualization methods must provide the user access to uncertainty in classification or image generation if the results of the visualization are to be trusted. This work makes the concept of uncertainty a central component of visualization. Users should be able to assess uncertainty, and therefore potential for error, while exploring the space of possibility with respect to concrete realizations of ambiguous features. This research achieves the following: (1) development of new visualization methods that expose uncertainty to users in a meaningful way, (2) adaptation of advanced classification algorithms so that partial class probabilities, and therefore uncertainty, are preserved, and (3) development a rigorous validation system based on generated ground truth models and simulated acquisition. The semantic meaning of uncertainty is a guiding principle of this work. This means that when data are processed and classified, a range of possible variations are realizable as a consequence of uncertainty. Contrast this with a syntactic approach, where descriptions such as fuzziness and noise are taken literally, and are used as visual metaphors to indicate uncertainty. This research shows that semantic methods of image representation and interaction are more amenable to expert analysis and hypothesis testing.