Title: Designing, Developing, and Democratizing Guidance for Visual Analytics
Date: Tuesday, November 19, 2024
Time: 8–10 AM Eastern Time (US)
Location: TSRB 334 (VIS Lab) – just walk in, show your BuzzCard to the concierge if asked
Virtual Meeting: Zoom
Arpit Narechania
https://arpitnarechania.github.io
Computer Science PhD Student
School of Interactive Computing
Georgia Institute of Technology
Committee:
Dr. Alex Endert – Advisor, Georgia Tech, School of Interactive Computing
Dr. John Stasko – Georgia Tech, School of Interactive Computing
Dr. Duen Horng (Polo) Chau – Georgia Tech, School of Computational Science & Engineering
Dr. Clio Andris – Georgia Tech, School of City and Regional Planning
Dr. Shamkant B. Navathe – Georgia Tech, School of Computer Science
Dr. Mennatallah El-Assady – ETH Zürich, ETH AI Center
Abstract:
The ubiquity and utility of data today underscore the importance of balancing system efficiency with human intuition and expertise to ensure timely and accurate decision-making. Yet, challenges arise when users must extensively input their analytic intent or when automated system actions misinterpret their needs. “Guidance” – or any kind of help, tip, advice, support, suggestion, or recommendation – offers a promising solution to bridge this “knowledge gap” between the two, enhancing both the quality of analysis and making the process more enjoyable for users. This thesis extends information visualization (InfoVis), visual analytics (VA) and human-computer interaction (HCI) literature by contributing a series of mixed-initiative guidance-enriched tools and techniques, wherein the user and the system both learn from and take initiative on behalf of each other to steer the analysis process. We categorize these contributions under three main thrusts:
(1) Investigate the Role of Guidance in Visual Analytics: We introduce a data preparation system (DataPilot) that utilizes data quality and usage insights to guide users in selecting effective subsets from large, unfamiliar tabular datasets. Bringing the human into the (analysis) loop, we introduce a question-answering system integrated with a self-service debugging view (DIY – Debug-It-Yourself), that helps users interactively assess the correctness of natural language to SQL workflows.
(2) Develop Mixed-Initiative Guidance Systems. We introduce a visual data analysis system (Lumos) that increases users’ awareness of (biased) analytic behaviors by comparing it against a target behavior; users can adjust this target behavior, achieving co-adaptivity. We enhance Lumos into a multimodal data analysis system (BiasBuzz) that additionally provides haptic feedback to quickly draw users’ attention in case of significantly biased behaviors. As a culmination system, we introduce the first mixed-initiative visual data analysis system that can seamlessly transition between different levels of guidance based on the analysis needs and user preferences (Lighthouse).
(3) Democratize Building Custom Guidance Systems. We contribute an open-source library of enhanced user interface (UI) controls that track and dynamically overlay analytic provenance, enabling developers to prototype custom guidance-enriched systems (ProvenanceWidgets). To ensure consistent and effective user interfaces for visual data analysis, we also derive two design spaces: one for communicating analytic provenance (ProvenanceLens) and another for communicating guidance (Lighthouse).
The outcomes of this thesis have been disseminated through multiple publications in top journals and conferences (TVCG, VIS, CHI, IUI), multiple patent filings by Adobe and Microsoft, integration into an Adobe product, open-source software, and inclusion in coursework on visualization and human-centered data analysis at Georgia Tech, achieving broader impact for researchers, developers, practitioners, and students alike.