CHS: Small: Collaborative Research: Representing and Learning Visualization Design Knowledge

To create effective visualizations, designers must consider both the data domain under consideration and principles of effective visual encoding. Both classic and recent research attempts to formalize such design knowledge by representing visualizations as a collection of logical facts and guidelines as logical constraints over these facts. Given such a knowledge base, constraint solvers can be used to automatically critique and synthesize visualization designs, even given only partial specifications as input. However, though the rule sets that constitute these knowledge bases are informed by perceptual research, they are engineered by hand, limited in scope to static single-view displays, and often biased towards the needs of the specific tool for which they were crafted. To enable more powerful automated design tools and develop a shared community resource of actionable design knowledge, we propose to investigate new methods for representing and learning visualization design guidelines encoded as declarative constraints. First, we will research fundamental experimental methods, and user interface design issues, and statistical models for identifying, aggregating, browsing, editing, and testing visualization knowledge bases. KA key components of this effort areis the development of methods for reliably identifying principles using existing graphical perception and cognition literature, and design of experimental protocols for eliciting guidelines from visualization experts or practitioners, and for collecting task-specific perceptual judgments for automatically learning appropriate weights for soft constraints. Second, we will use these techniques to develop novel extensions to visualization design knowledge bases. By incorporating descriptive statistics of input datasets as input "facts", we plan to develop design guidelines for automatically adapting standard visualization designs to accommodate large datasets. Next, we will develop an integrated knowledge base for supporting the design of multi-view displays (such as dashboards) that must balance concerns of the effectiveness of individual views with concerns for global ordering and consistency. For each of these application domains, we will develop novel logical representations of visualization design guidelines, collect experimental data to resolve and appropriately weight conflicts among these rules, and evaluate the resulting systems through human-subjects performance studies.