Date of Award
Summer 9-15-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
We are now able to collect data from most aspects of our lives on Earth and beyond. The abundance of data, however, is not equivalent to an abundance of actionable knowledge. The data needs to be effectively analyzed and transformed into knowledge. Visual analytics, as a scientific discipline, aims to study how humans and computers unify their respective strengths to reason about large amounts of data effectively. With recent advances in artificial intelligence, the role of computers in this symbiosis has evolved, creating new research frontiers to understand how we can create human-computer partnerships for visual data analysis. Furthermore, given that users are different and analyze datasets differently, building machines that effectively understand and assist users is non-trivial. In this dissertation, we make three primary contributions. First, we propose an agent-based conceptual framework to formalize mixed-initiative visual analytics, where human agents and artificial agents interact with their analytic environment to effectively analyze data and make decisions. In light of this framework, we demonstrate how we can reason about existing visual analytics research, identify gaps for future research, and design new visual analytic systems for domain experts. Second, we consider how computers learn from users by observing their low-level interactions with the visual interface. In doing so, we utilize the Bayesian model selection framework. In our approach, we enumerate a set of possible exploration patterns as probabilistic models and maintain a posterior belief over the viability of each model in light of observed user interactions. Validating our approach on three existing user studies, our results demonstrate that we can effectively detect exploration biases and predict future data interactions. Our approach is not the only one in the literature. Given that there are other similar techniques, we conduct a study to compare and contrast a set of related user modeling techniques on several different interaction log datasets. To go beyond predicting bias and data interaction, we also consider how to collapse interaction logs into more meaningful tokens that describe high-level tasks or reasoning processes. In doing so, we utilize a grammar-based method to apply existing task taxonomies to interaction logs. Third, we investigate a mixed-initiative system that assists users in data exploration and discovery -- an important phase of the sense-making process known as data foraging. We select an active search algorithm (a flavor of Bayesian optimization) to be the artificial agent in our system. In this setting, the system observes user interactions and infers their latent interests. Then, the active search algorithm manipulates the visual interface by highlighting the most promising points as candidates for future interactions. The objective of the active search algorithm is to maximize the discovery throughput. We conduct a crowd-sourced user study on this system and demonstrate its impact on discovery throughput as well as human factors including trust, engagement, and biases. We conclude this dissertation by providing an in-depth discussion of the implications of our findings in building the next generation of intelligent visual analytic systems. Specifically, we will discuss our vision and some ongoing work for such systems catered toward experts in the material science domain.
Language
English (en)
Chair
Alvitta Ottley
Committee Members
Roman Garnett