Tamara Munzner

Prototyping is employed in information visualization to understand user needs and to iteratively implement, test, and refine possible solutions. Yet, data is often seen merely as a resource in this process, and toy or synthetic datasets can lead to incorrect data abstractions and less effective visualization designs.In this dissertation, we demonstrate how a data-driven prototyping process based on real-world data can lead to novel contributions in information visualization with high industry relevance.The design study on the Ocupado project describes the process of designing, implementing, and evaluating a suite of novel visualization tools for studying space utilization at scale. We reflect on the prototyping process that included multiple stakeholders and present generalizable design choices for visualizing non-trajectory spatiotemporal data related to indoor regions.The findings from the Ocupado study highlight the need for analyzing data concerning time periods of interest that are known in advance rather than determined on the fly. We provide a detailed characterization of non-contiguous time series slices and propose TimeElide, a domain-agnostic visual analysis tool and design space.Inspired by emerging large-scale visualization collections and the difficulty in finding relevant information, we investigate a content-based approach for visualization recommendation in the VizCommender study. We focus on text-based content that is representative of the subject matter of visualizations and compare different similarity measures.We identify that all existing visualization snippets—compact previews of visualizations in those collections—are characterized by their low information density and fail to help people judge the relevance. The VizSnippets study is the first systematic approach to visualization snippet design. We propose a design framework and computational pipeline for the lossy compression of visual and textual content into representative snippets.A critical reflection on our data-driven prototyping approach, and visualization design studies in general, reveals an alternative avenue for applied visualization projects that begins with real-world data rather than specific stakeholder analysis questions. We introduce the notion of data-first design studies and provide practical guidance.

View record

Visualization is an important tool to analyze data, but there emerge various challenges from complex analytical data and tasks. In this dissertation, I present four projects that were motivated by these challenges, situated in the nested model proposed by Munzner, which consists of four layers to describe the components in visualization: domain, data and task, encoding design, and algorithm.In ADVIEW, to address the challenges of comparing many phylogenetic trees in the domain of biology, I propose a visual encoding to compress a tree representation, design and implement a multi-view interactive tool to handle the multiple levels of detail in a tree collection dataset, ranging from the whole collection, through subsets of trees, individual trees, subtrees, to leaf nodes.In SPRAWLTER, to address the existing visual encoding problems of readability metrics for node-link graphs, I propose two novel metrics to measure a finer-grained clutter and to balance the geometric sparseness and clutter. These metrics recognize different levels of visual saliency such as metanodes and leaf nodes in multi-level graphs. In LOGSEG, to fulfill user demands for chunking actions in the domain of image editing software, I propose a segmentation model for the action logs to serve the demands that require different chunking granularities. For example, smart undo for going back to a previous user task needs a low-level chunking, while managing an overview of milestones needs a high-level one.In CORGIE, to fill the gap in visual qualitative evaluation of graph neural networks (GNNs) in the domain of machine learning, I propose an approach and design a tool to explore correspondences between a graph and its embedding to check how different levels of structures are preserved from the input graph to the output embedding. I also design a new graph layout to reveal how a GNN leverages node neighbors and computes an embedding.I identify a common theme among these projects: multi-level structures. They consist of nesting subsets of data points that are relevant to the analytical tasks. I demonstrate how to exploit them in the visualization if provided in hierarchical data, or to compute them for non-hierarchical data.

View record

Technological innovations have allowed for a greater variety of data, most notably microbial genomic data, to be collected, integrated, analyzed, and visualized for epidemiological investigations. While analytic methods have evolved in light of this technological change, data visualizations systems have lagged behind.I take a novel approach that integrates methods from information visualization, human computer interaction, machine learning, and statistics to address unmet data visualization needs in microbial genomic epidemiology (genEpi). This approach also enables me to generate study artifacts that can be used to address regulatory and organizational constraints arising in domains where the use of data is highly restricted. I first present a mixed methods approach to understand the needs, data, tasks, and constraints of public health stakeholders that are charged with interpreting the findings of these data. I demonstrate how this approach can be used to communicate new and heterogeneous types of data in a clinical report that is read by stakeholders in different roles. I next present a novel method for systematically reviewing data visualizations that I use to develop a Genomic Epidemiology Visualization Typology (GEViT), which enables others to explore and characterize the way the data could be visualized. Finally, I use these collective findings to inform the design and implementation of data visualization tools: Adjutant, the GEViT Gallery, minCombinR, and GEViTRec. Adjutant enables rapid and unsupervised topic clustering of PubMed article corpuses to aid systematic and literature reviews. The GEViT gallery is a browsable interface for exploring data visualizations specific to the microbial genEpi domain. minCombinR lowers the burden to stakeholders for generating combinations of data visualizations for heterogeneous data. Finally, GEViTRec takes a novel approach to the automatic generation of data visualizations that can help stakeholders familiarize themselves with new data. All of these tools integrate with analytic methods.This research makes novel contributions to the design and implementation of data visualization systems that impact microbial genomic epidemiological data collected for public health investigations. The challenges addressed here are not unique to this domain and my contributions are extensible to other domains grappling with heterogeneous, multidimensional, and restricted data.

View record

Why do people visualize data? People visualize data either to consume or produce information relevant to a domain-specific problem or interest. Visualization design and evaluation involves a mapping between domain problems or interests and appropriate visual encoding and interaction design choices. This mapping translates a domain-specific situation into abstract visualization tasks, which allows for succinct descriptions of tasks and task sequences in terms of why data is visualized, what dependencies a task might have in terms of input and output, and how the task is supported in terms of visual encoding and interaction design choices. Describing tasks in this way facilitates the comparison and cross-pollination of visualization design choices across application domains; the mapping also applies in reverse, whenever visualization researchers aim to contextualize novel visualization techniques. In this dissertation, we present multiple instances of visualization task abstraction, each integrating our proposed typology of abstract visualization tasks. We apply this typology as an analysis tool in an interview study of individuals who visualize dimensionally reduced data in different application domains, in a post-deployment field study evaluation of a visual analysis tool in the domain of investigative journalism, and in a visualization design study in the domain of energy management. In the interview study, we draw upon and demonstrate the descriptive power of our typology to classify five task sequences relating to visualizing dimensionally reduced data. This classification is intended to inform the design of new tools and techniques for visualizing this form of data. In the field study, we draw upon and demonstrate the descriptive and evaluative power of our typology to evaluate Overview, a visualization tool for investigating large text document collections. After analyzing its adoption by investigative journalists, we characterize two abstract tasks relating to document mining and present seven lessons relating to the design of visualization tools for document data. In the design study, we demonstrate the descriptive, evaluative, and generative power of our typology and identify matches and mismatches between visualization design choices and three abstract tasks relating to time series data. Finally, we reflect upon the impact of our task typology.

View record

In this thesis, we explore ways to make practical extensions to Dimensionality Reduction, or DR algorithms with the goal of addressing challenging, real-world cases. The first case we consider is that of how to provide guidance to those users employing DR methods in their data analysis. We specifically target users who are not experts in the mathematical concepts behind DR algorithms. We first identify two levels of guidance: global and local. Global user guidance helps non-experts select and arrange a sequence of analysis algorithms. Local user guidance helps users select appropriate algorithm parameter choices and interpret algorithm output. We then present a software system, DimStiller, that incorporates both types of guidance, validating it on several use-cases. The second case we consider is that of using DR to analyze datasets consisting of documents. In order to modify DR algorithms to handle document datasets effectively, we first analyze the geometric structure of document datasets. Our analysis describes the ways document datasets differ from other kinds of datasets. We then leverage these geometric properties for speed and quality by incorporating ideas from text querying into DR and other algorithms for data analysis. We then present the Overview prototype, a proof-of-concept document analysis system. Overview synthesizes both the goals of designing systems for data analysts who are DR novices, and performing DR on document data. The third case we consider is that of costly distance functions, or when the method used to derive the true proximity between two data points is computationally expensive. Using standard approaches to DR in this important use-case can result in either unnecessarily protracted runtimes or long periods of user monitoring. To address the case of costly distances, we develop an algorithm framework, Glint, which efficiently manages the number of distance function calculations for the Multidimensional Scaling class of DR algorithms. We then show that Glint implementations of Multidimensional Scaling algorithms achieve substantial speed improvements or remove the need for human monitoring.

View record

A graph consists of a set and a binary relation on that set. Each elementof the set is a node of the graph, while each element of the binary relationis an edge of the graph that encodes a relationship between two nodes.Graph are pervasive in many areas of science, engineering, and the socialsciences: servers on the Internet are connected, proteins interact in largebiological systems, social networks encode the relationships between people,and functions call each other in a program. In these domains, the graphscan become very large, consisting of hundreds of thousands of nodes andmillions of edges.Graph drawing approaches endeavour to place these nodes in two orthree-dimensional space with the intention of fostering an understandingof the binary relation by a human being examining the image. However,many of these approaches to drawing do not exploit higher-level structuresin the graph beyond the nodes and edges. Frequently, these structures canbe exploited for drawing. As an example, consider a large computer networkwhere nodes are servers and edges are connections between those servers.If a user would like understand how servers at UBC connect to the rest ofthe network, a drawing that accentuates the set of nodes representing thoseservers may be more helpful than an approach where all nodes are drawn inthe same way. In a feature-based approach, features are subgraphs exploitedfor the purposes of drawing. We endeavour to depict not only the binaryrelation, but the high-level relationships between features.This thesis extensively explores a feature-based approach to graph visualization and demonstrates the viability of tools that aid in the visualization of large graphs. Our contributions lie in presenting and evaluatingnovel techniques and algorithms for graph visualization. We implement fivesystems in order to empirically evaluate these techniques and algorithms,comparing them to previous approaches.

View record

Large data sets are difficult to analyze. Visualization has been proposed to assist exploratory data analysis (EDA) as our visual systems can process signals inparallel to quickly detect patterns. Nonetheless, designing an effective visualanalytic tool remains a challenge.This challenge is partly due to our incomplete understanding of how commonvisualization techniques are used by human operators during analyses, either inlaboratory settings or in the workplace.This thesis aims to further understand how visualizations can be used to support EDA. More specifically, we studied techniques that display multiple levels of visual information resolutions (VIRs) for analyses using a range of methods.The first study is a summary synthesis conducted to obtain a snapshot ofknowledge in multiple-VIR use and to identify research questions for the thesis:(1) low-VIR use and creation; (2) spatial arrangements of VIRs. The next twostudies are laboratory studies to investigate the visual memory cost of imagetransformations frequently used to create low-VIR displays and overview usewith single-level data displayed in multiple-VIR interfaces.For a more well-rounded evaluation, we needed to study these techniques inecologically-valid settings. We therefore selected the application domain of websession log analysis and applied our knowledge from our first three evaluationsto build a tool called Session Viewer. Taking the multiple coordinated viewand overview + detail approaches, Session Viewer displays multiple levels ofweb session log data and multiple views of session populations to facilitate dataanalysis from the high-level statistical to the low-level detailed session analysisapproaches.Our fourth and last study for this thesis is a field evaluation conducted atGoogle Inc. with seven session analysts using Session Viewer to analyze theirown data with their own tasks. Study observations suggested that displayingweb session logs at multiple levels using the overview + detail technique helped bridge between high-level statistical and low-level detailed session analyses, andthe simultaneous display of multiple session populations at all data levels usingmultiple views allowed quick comparisons between session populations. We alsoidentified design and deployment considerations to meet the needs of diversedata sources and analysis styles.

View record