On the optimization of visualizations of complex phenomena

The problem of perceptually optimizing complex visualizations is a difficult one, involving perceptual as well as aesthetic issues. In our experience, controlled experiments are quite limited in their ability to uncover interrelationships among visualization parameters, and thus may not be the most useful way to develop rules-of-thumb or theory to guide the production of high-quality visualizations. In this paper, we propose a new experimental approach to optimizing visualization quality that integrates some of the strong points of controlled experiments with methods more suited to investigating complex highly-coupled phenomena. We use human-in-the-loop experiments to search through visualization parameter space, generating large databases of rated visualization solutions. This is followed by data mining to extract results such as exemplar visualizations, guidelines for producing visualizations, and hypotheses about strategies leading to strong visualizations. The approach can easily address both perceptual and aesthetic concerns, and can handle complex parameter interactions. We suggest a genetic algorithm as a valuable way of guiding the human-in-the-loop search through visualization parameter space. We describe our methods for using clustering, histogramming, principal component analysis, and neural networks for data mining. The experimental approach is illustrated with a study of the problem of optimal texturing for viewing layered surfaces so that both surfaces are maximally observable.