A Visual Analytic Framework for Graphs.

Abstract

Visual Analytics is the science of analytical reasoning through visual interaction. Graphs are ubiquitous: high-throughput ``omics'' sciences generate data to study pathways and computer networks use graphs to analyze communications. Graphs are also particularly amenable to visual representation. It is no surprise that graph visual analytics has received a lot of attention as evidenced by the large number of tools and algorithms developed for this purpose. These systems can be very informative, but usually constrain the reader in realizing their full value for the following reasons: they are unable to be scaled or modified for changing and increasing data and they do not support missing attributes and uncertainty in datasets. This dissertation presents several components within a visual analytic framework to address these problems.

To support scalable and modifiable methods for graph visual analytics, we introduce a visual analytic graph algebra and its atomic operators, which include selection and aggregation. We demonstrate an implementation of the algebra through a plugin on a widely used visualization tool, Cytoscape. We then use Cytoscape and the algebra to create visualizations to replicate previous statistical studies on high-throughput biological datasets. Next, we present optimization techniques to accelerate the visual exploration and discovery process for repeat workflows. Within our framework, we present a multi-modal graph exploration tool, GreenTrellis to address interaction challenges with large graphs. GreenTrellis links multiple graph signature scatterplots to characterize the graph at different node localities. We present several rich signature representations based on different characteristics and show how they can be used in different analytic tasks through example and user study. Interacting with real world data is messy, data collection often results in missing data; furthermore, the user may have a degree of uncertainty about a desired query. We develop a pattern-matching based algorithm to predict values for missing attributes at nodes. Finally, we introduce a probabilistic framework within the visual analytic algebra that incorporates uncertainty in the queries and provides a probabilistic assessment of the likelihood of the final obtained outcomes.