Representing Visualization Insights as a Dense Insight Network

TL;DR

This work introduces a dense insight network to encode and explore relationships among automatically generated dashboard insights, organized across five link categories: type-based, topic-based, value-based, metadata-based, and score-based. A visualization playground demonstrates how these connections can be interacted with, while a case study with GPT-3.5 shows how the network supports ranking, ordering, and prompting for concise LLM-based dashboard summaries. The framework enables controlled, explainable insight selection, reduces LLM hallucination risk, and provides a foundation for designing new analytics tools and evaluation methods. Overall, the approach advances interpretable, user-guided narrative generation for complex dashboards, with practical implications for automated storytelling and dashboard communication.

Abstract

We propose a dense insight network framework to encode the relationships between automatically generated insights from a complex dashboard based on their shared characteristics. Our insight network framework includes five high-level categories of relationships (e.g., type, topic, value, metadata, and compound scores). The goal of this insight network framework is to provide a foundation for implementing new insight interpretation and exploration strategies, including both user-driven and automated approaches. To illustrate the complexity and flexibility of our framework, we first describe a visualization playground to directly visualize key network characteristics; this playground also demonstrates potential interactive capabilities for decomposing the dense insight network. Then, we discuss a case study application for ranking insights based on the underlying network characteristics captured by our framework, before prompting a large language model to generate a concise, natural language summary. Finally, we reflect on next steps for leveraging our insight network framework to design and evaluate new systems.

Figures (4)

• Figure 1: An example dashboard titled "Call Center Overview" (left) and the same dashboard with the underlying data tables shown for all sub-panels instead of the visualizations (right). The dashboard has five sub-panels showing the average duration and number of calls to a call center, broken down by the sentiment and reason for the call: A a line chart showing the total number of Calls per Date; B a bar chart of the total number of Calls for each Day of the week; C a donut chart of the number of Calls by Sentiment; D a data table showing the number of calls by Sentiment, as well as the average call Duration broken down by both call Reason and Sentiment; and E a multi-line chart showing the number of Calls by Sentiment per Date. For demonstration purposes, we generate forty-nine insights for this sample dashboard (Section \ref{['sec:overview']}). Note that the dashboard colors were chosen for illustrative purposes to match the network figures in this paper. The dashboard description shows the resulting LLM-based summary produced via our case study application (Section \ref{['sec:llm']}).
• Figure 2: Sample insight clusters produced in our visualization playground based on different types of links in our insight network framework. Each cluster is a clique, with links connecting every pair of nodes in the cluster. Each sub-figure only includes one node per insight unless otherwise indicated; the clustering for "Dimensions" (G) is the notable exception, as some insights correspond to multiple dimensions.
• Figure 3: Our visualization prototype has two components: (A) the interactive visualization panel supports exploration of linked insights using different visual representations (such as the node-link network view shown here); (B) the story exploration panel includes (c) a concatenated insight paragraph and (d) linear narrative components containing the insights and accompanying visualizations.
• Figure 4: (A) The matrix visualization and (B) cluster visualization of the priority score for the subest of seven insights in Figure \ref{['fig:teaser']}.