Publication Type

PhD Dissertation

Version

publishedVersion

Publication Date

7-2026

Abstract

Automated visualization systems reduce the expertise required to create charts by generating visual representations directly from input data. Despite substantial advances in recommendation accuracy and generation quality, these systems share a common and largely unexamined assumption: that the pipeline from input to output is usage agnostic, and that the situation in which a system is deployed does not impose design constraints that should shape its output. This dissertation argues that this assumption is wrong, and addresses the following research question: how can we characterize the design constraints imposed by distinct usage scenarios, and incorporate those constraints into automated visualization generation?
We demonstrate this across three usage scenarios, each representing a different input modality, user population, and physical context. The first scenario, layperson analytical usage, arises when users without a visualization background rely on automated recommendation systems to analyze tabular data. We present AdaVis, a system that reframes recommendation as a one to many mapping problem using a knowledge graph with box embeddings, and generates natural language explanations for each candidate chart. Quantitative evaluations and user interviews demonstrate that AdaVis successfully recommends multiple valid visualizations and produces explanations judged correct and useful by both novice and expert users. The second scenario, public mobile display usage, arises when people view sensitive personal data visualizations on smartphones in public settings where bystanders can observe the screen. We present a perception driven system that adjusts the spatial frequency and luminance contrast of visual marks so that a chart remains legible at close proximity while becoming uninterpretable at a greater distance. Two user studies with 16 and 18 participants respectively demonstrate that the system achieves comparable legibility to unmodified visualizations at 30 centimeters while substantially reducing interpretability at 90 centimeters. The third scenario, text based usage, arises when visualizations are generated automatically from natural language documents that contain linguistic uncertainty markers. We present UncertaintyVis, a system that classifies linguistic uncertainty into four categories derived from a formative corpus analysis and maps each category to chart specific visual encodings. A two part user study with 12 participants demonstrates that chart to text matching achieves 85% accuracy, that uncertainty aware visualizations reduce perceived cognitive demand during document reading, and that 75% of participants prefer uncertainty preserving charts over plain text.
Together, these three systems establish that identifying usage scenarios and incorporating their design constraints into the generation process produces outputs that generic pipelines cannot achieve. Usage scenario awareness is a necessary and productive design principle for the next generation of automated visualization research.

Degree Awarded

PhD in Computer Science

Discipline

Databases and Information Systems | Graphics and Human Computer Interfaces

Supervisor(s)

TANG, Anthony Hoi Tin

First Page

1

Last Page

232

Publisher

Singapore Management University

City or Country

Singapore

Copyright Owner and License

Author

Download

Available for download on Thursday, August 26, 2027

Share

COinS