VisAider: AI-Assisted Context-Aware Visualization Support for Data Presentations
Kentaro Takahira, Yuki Ueno
TL;DR
The paper tackles the challenge of real-time data presentation in small-group settings, where audience interests and questions require dynamic visualization adjustments beyond fixed action mappings. It proposes VisAider, an AI-assisted agent that analyzes four contextual inputs—the available dataset, the active visualization, ongoing conversation, and audience profile—to generate ranked Vega-Lite visualization candidates via a LangChain-based pipeline powered by GPT-4o. Grounded in a formative study with experienced analysts, the work identifies four design considerations: preserve presenter autonomy, enable low-effort integration, ensure relevance to current content, and tailor visualization formats to the audience. Preliminary prototypes in simulated scenarios show feasibility but reveal challenges in inferring complex data transformations, resolving ambiguous tasks, and achieving low latency, guiding planned optimizations and an upcoming summative user study. Overall, the work provides a concrete framework and prototype toward fluid, context-aware data communication in live presentations, with potential to enhance presenters' ability to adapt visuals in real time toward audience needs.
Abstract
Effective real-time data presentation is essential in small-group interactive contexts, where discussions evolve dynamically and presenters must adapt visualizations to shifting audience interests. However, most existing interactive visualization systems rely on fixed mappings between user actions and visualization commands, limiting their ability to support richer operations such as changing visualization types, adjusting data transformations, or incorporating additional datasets on the fly during live presentations. This work-in-progress paper presents VisAider, an AI-assisted interactive data presentation prototype that continuously analyzes the live presentation context, including the available dataset, active visualization, ongoing conversation, and audience profile, to generate ranked suggestions for relevant visualization aids. Grounded in a formative study with experienced data analysts, we identified key challenges in adapting visual content in real time and distilled design considerations to guide system development. A prototype implementation demonstrates the feasibility of this approach in simulated scenarios, and preliminary testing highlights challenges in inferring appropriate data transformations, resolving ambiguous visualization tasks, and achieving low-latency responsiveness. Ongoing work focuses on addressing these limitations, integrating the system into presentation environments, and preparing a summative user study to evaluate usability and communicative impact.