Visualizing Tree-of-analysis: Facilitating Conversational Visual Analytics for Novices

Paper Title

Visualizing Tree-of-Analysis: Facilitating Conversational Visual Analytics for Novices

Publication Info

• Topic area: Conversational visual analytics (CVA) for novice users
• Keywords: Conversational visual analytics, tree visualization, novice users, large language models, data exploration, user interface, query recommendation, analysis workflow, multimodal interaction, democratizing analytics

Background and Problem

• Problem / challenge: Novices using conversational visual analytics (CVA) often become disoriented due to fragmented analysis flows, vague queries, and insensitivity to analytical cues in AI outputs. Existing solutions provide isolated recommendations but fail to offer a global perspective of the analysis journey.
• Significance: Addressing disorientation in CVA is critical for making data exploration accessible to novices, enabling them to perform meaningful analysis without requiring advanced expertise in data science or visualization tools.
• Motivation and related work: Prior research in CVA, visual analytics, and conversational systems has focused on improving input clarity, generating follow-up questions, and providing data-centric recommendations. However, these approaches lack mechanisms to help users maintain a holistic view of their analysis process. Inspired by tree structures used in complex task visualization, this paper seeks to bridge this gap by structuring CVA interactions into an analysis tree.

Solution

• Proposed approach: Tree-of-Analysis (ToA), a system that organizes CVA conversations into an interactive analysis tree, with AI outputs as nodes and user queries as categorized edges.
• Novelty:
  1. Introduction of an analysis tree structure to provide a global overview of CVA processes.
  2. Development of a real-time tree construction algorithm that extracts analytical cues and categorizes user queries into seven types.
  3. Integration of tree-based query recommendations to guide novices in their exploration.
  4. Validation through user studies and expert interviews, demonstrating improved task performance and reduced disorientation.
• Procedure and key techniques:
  1. Conducted a formative study to identify novice weaknesses (e.g., cue insensitivity, vague queries) and expert strengths (e.g., trigger sensitivity, tree-like workflows).
  2. Designed a tree structure where AI outputs (charts, text, triggers) form nodes, and user queries (categorized into seven types) form edges.
  3. Developed an LLM-based algorithm to construct the tree in real-time by detecting triggers and connecting nodes based on predefined rules.
  4. Enhanced the CVA interface with a tree navigation panel and contextual query recommendations.
  5. Evaluated ToA through a user study (N = 12 novices) and expert interviews (N = 3).

Results

• Concrete findings:
  • ToA increased per-turn insights by 58.3% (1.167 insights/turn vs. 0.737 with baseline).
  • ToA required 17.7% longer per-turn thinking time (71.3 seconds vs. 60.6 seconds with baseline).
  • ToA reduced conversation turns by 23.1% (14.4 turns vs. 18.8 with baseline).
  • All participants successfully completed tasks with ToA, while one failed with the baseline system.
• Advantage over baselines:
  • ToA eliminated task failures and improved insight productivity.
  • Reduced mental demand, effort, and improved perceived performance (NASA-TLX scores).
  • Enhanced user perception of usefulness, particularly in tracking analysis progress and identifying relationships between steps.
• Experiments / evaluation:
  • User study (N = 12 novices) comparing ToA with a baseline CVA system without the analysis tree.
  • Technical evaluation of tree construction accuracy (70.73% overall, with 99.10% accuracy in query type classification).
  • Expert interviews (N = 3) assessing ToA’s expressiveness, usability, and pedagogical value.
• Limitations and future work:
  • Tree complexity may become overwhelming in extended analyses.
  • Rule-based tree construction may misalign with user logic in complex scenarios.
  • Limited evaluation scope (two datasets, short-term sessions, small sample size).
  • Future work includes adding interactive pruning, supporting storytelling, and expanding ToA to more analytical tasks.

Summary

This paper introduces Tree-of-Analysis (ToA), a novel system that structures conversational visual analytics (CVA) into an interactive analysis tree to help novices maintain orientation during data exploration. By categorizing user queries and visualizing AI outputs as nodes, ToA provides a global overview of the analysis process and offers contextual query recommendations. User studies demonstrate that ToA significantly improves insight productivity (+58.3%) and eliminates task failures, albeit with slightly longer per-turn thinking times. Expert interviews highlight its potential for democratizing visual analytics and teaching data exploration skills. Future work will address scalability, storytelling applications, and broader task support.