Accelerating the Improvement of Visualization Design Methodology

点击翻译按钮获取中文摘要

My research creates interactive visualization tools to help people find patterns in data. For this grant I propose work that will lead to a faster and better process for making these tools. The modern era is characterized by the promise of better decision making through access to more data than ever before. When people have well-defined questions to ask about data, they can use purely computational techniques from fields such as statistics and machine learning. However, many analysis problems are ill specified: people don't know which of many questions to ask. In such cases, the best path forward is an analysis process with a human in the loop, exploiting the powerful pattern detection properties of the human visual system. People use visualization ("vis") tools for many reasons. They can be stepping stones to gaining a crisper understanding of analysis requirements before developing formal models. They might be aimed at the designers of another system to help them refine that system's algorithms, or at end users of another system who are trying to decide its behavior is reasonable, for example to see if the results of a machine learning system are trustworthy before deploying it. In addition to these transitional uses, vis tools are also designed for permanent use, where a human intends to stay in the loop indefinitely. A common case is exploratory analysis for scientific discovery, where the goal is to speed up and improve a human's ability to generate and check hypotheses. An example from my own work is a tool to help biologists studying the genetic basis of disease through analyzing DNA sequence variation. Vis tools can also be used for presentation; for example, the New York Times has deployed sophisticated interactive visualizations. Many vis idioms have been proposed: loosely speaking, idioms are distinct approaches to creating and manipulating visual representations, ranging from simple scatterplots to complex depictions of network structure at different scales. However, end users are offered little general guidance about how to choose an appropriate idiom for their domain. The research I propose here will bridge this gap, offering a general methodology for vis design. This will allow vis practitioners to develop better tools more quickly, thus providing benefits to the many people who rely upon vis for business-critical decisions. I am well positioned to conduct the proposed work because of my experience in three other threads of research, all of which I am continuing to advance under other funding. One is also methodological, but addresses the high-level process of vis research by proposing research models. The other two, the “technique-driven” and “problem-driven” legs of my research program, involve building tools. I will use three main evaluation methods in my new work on methodologies for vis design. I will conduct (1) controlled quantitative lab studies and (2) qualitative field studies, both of which are techniques with a long history in human-computer interaction and cognitive psychology. I have also proposed a new method, (3) data studies, which invert the emphasis of lab studies by collecting judgements about a large number of datasets from a small number of human experts. A major focus of my research program is providing HQP with vis expertise to Canadian industry and academia; the demand already outstrips the supply. The need for vis tools is becoming increasingly pervasive across industry, science, health, and the exploration of personal and public data. Supporting foundational visualization research in Canada also opens the door to the creation of spinoff companies; for example, a recent US vis startup (Tableau) was one of the biggest IPOs of 2013, and now has a $4B market capitalization.