PaperlayerView exact revision on arXiv

Source-linked AI summary

Living Dashboards: Automatically Self-Updating Visualization Dashboards

Mingyu An, Heyon Jeon, Sungbok Shin, Jinwook Seo, Eduard Gröller, Niklas Elmqvist, Vaishali Dhanoa

arXiv:2610.08393v1cs.HC

TL;DR

Dashboards can update their data while their fixed structure may not keep pace with users’ changing analytical questions. Living Dashboards adapt views in response to observed use and reserve consequential changes for user approval. In an exploratory study, participants answered more tasks correctly and rated usability higher than an AI-supported baseline, though the comparison did not isolate adaptation alone.

  • Problem

    Dashboards’ fixed structures may struggle to keep pace with users’ evolving analytical questions even as their data updates.

    Source: §1 · p. 2

  • Method

    Living Dashboards use a focus score derived from observed interactions to wither or revive views autonomously while deferring additions and retirements to user approval.

    Source: §4 · p. 5, §5.2 · p. 9, §8 · p. 18

  • Results

    Living Dashboard participants averaged 9.17 correctly answered tasks (SD=1.17, range=7–10), versus 7.17 (SD=1.72, range=5–10) in Baseline; usability was higher, but conditions differed in more than adaptation alone.

    Source: §1 · p. 2, §6.5.2 · p. 14

  • Takeaways & Limitations

    The prototype demonstrates a dashboard-view lifecycle that responds to observed use while leaving consequential view changes to users.

    Source: §4 · p. 5, §8 · p. 18

    The decay rate was accelerated for the study, likely amplifying frustration during comparison tasks relative to real deployments where inactivity is measured over hours or days.

    Source: §7.5 · p. 18

Abstract

from arXiv · show

Visualization dashboards are widely used interactive tools, but a disconnect exists between the dynamic data they display and their static structure. End-users cannot modify the dashboard to answer new questions. We introduce Living Dashboards, whose views are born, wither, revive, and die in response to how they are used. Rather than requiring manual reconfiguration, a living dashboard observes interaction and natural-language queries to autonomously wither neglected views and revive those used again. More consequential decisions, such as adding or retiring views, are deferred to the user. We formalize the concept as a four-dimensional design space and implement it in Living Dashboard, a web-based prototype. We evaluate it in an exploratory between-subjects study (N = 12) against an AI-supported baseline on analytical tasks. Living Dashboard participants answered more tasks correctly, reported lower workload, and rated the system higher on usability, though the two conditions differed in more than adaptive behavior alone.

1 Introduction

Living Dashboards respond to changing user needs by adapting views as they are used, while keeping consequential changes under user control. An exploratory study reported better task accuracy and usability than an AI-supported baseline, with a comparison that does not isolate adaptation alone.

1 Introduction

  • Dashboards can stream fresh data while their fixed layouts, views, and encodings struggle to keep pace with users’ evolving analytical questions.Practitioner accounts describe dashboards becoming irrelevant as their original questions grow stale, sometimes prompting organizations to create new dashboards instead of maintaining existing ones.Source: §1 · p. 2
  • Living Dashboards autonomously wither or revive views based on observed use, while adding or retiring views requires user approval.Views shrink in size and visibility when neglected and return to their original state when interaction resumes.Source: §1 · p. 2
  • The design formalizes constraints around preserving users’ spatial mental maps, making changes perceptible, and keeping every view accessible, alongside four dimensions: Observation, Adaptation, Temporality, and Agency.Interaction is only a proxy for attention, since a view may remain useful for comparison or occasional reference despite little direct interaction.Source: §1 · p. 2
  • In an exploratory study with N = 12, Living Dashboard participants answered more tasks correctly and rated usability higher than the AI-supported baseline, though conditions differed in more than adaptation alone.The comparison included six participants per condition; the paper cautions that the observed differences reflect the combined effect of adaptation and AI assistance.Source: §1 · p. 2
  • The work combines a formative study with N = 4, a web-based prototype, and an exploratory between-subjects evaluation with N = 12.The study examined task performance, usability, and responses to changing view prominence.Source: §1 · p. 2

2 Related Work

Prior work informs Living Dashboards through dashboard authoring and recommendation, adaptive interfaces, and research on attention and focus. This paper shifts attention from creating views to how their prominence and lifecycle respond to use.

2 Related Work

  • Related work spans dashboard authoring and recommendation, adaptive interfaces and mixed-initiative systems, and attention and focus in visualization.The paper draws on these three areas to motivate its view-lifecycle approach.Source: §2 · p. 2

2.1 Dashboard Authoring and Recommendation

  • Automated tools such as VizML [19], Draco, LADV, DMiner, and LightVA [42] support creating or recommending visualizations; Living Dashboards addresses what happens after views are created.Prior work also describes dashboards as author-created and deployed for users, with analytical usefulness potentially limited as questions evolve.Source: §2.1 · pp. 2–3

2.2 Adaptive Interfaces and Mixed-Initiative Interaction in Visualization Systems

  • Adaptive-interface research includes systems that infer intent or adapt interfaces, preserve spatial consistency during gradual change, and balance automated initiative with user control.Living Dashboards applies these ideas to view lifecycle, combining automatic prominence changes with user approval for additions and retirement.Source: §2.2 · p. 3

2.3 Attention and Focus in Visualization

  • Attention research includes gaze-based evaluation and gaze-aware visualizations [34]; this work instead uses mouse interactions as an attention proxy to adapt views over time.Withering and revival also draw on overview+detail, zooming, focus+context, and visual cues that balance detail and context.Source: §2.3 · pp. 3–4

3 Extracting Design Goals

A formative study of dashboard users and authors surfaced substantial maintenance effort, varied approaches to unused dashboards, and concerns about customized AI dashboards. These findings motivate goals for reuse, managing underused views, adaptation through use, and continuity.

3.1 Formative Study

  • A formative study with four university-recruited participants—one dashboard user and three authors—used approximately 30-minute online interviews to inform design goals.Interview insights were extracted from transcripts and discussed in a group meeting.Source: §3.1 · p. 4

3.1.1 Monitoring Dashboard Usage

  • Dashboard use was monitored inconsistently: one author estimated 70% of created dashboards were inactive, another estimated 20%, and a user kept half as a reference library.One author reviewed unused dashboards every two years, while another checked periodically or relied on users to report disuse.Source: §3.1.1 · p. 4

3.1.2 Understanding Dashboard Decay

  • Participants linked dashboard decay to quality, training, data issues, lack of awareness, mismatched or overwhelming views, and costly maintenance; responses ranged from replacement to archiving or refinement.A3 redirected maintenance effort toward new dashboards, while A4 combined feedback-based refinements with periodic reviews.Source: §3.1.2 · p. 4

3.1.3 Concerns Around Customized AI Dashboards

  • Participants raised AI-dashboard concerns including data security and GDPR, reduced standardization, uncertain quality and trust, difficulty extending outputs, and users’ limited time, skills, or data access.They also questioned whether users would create dashboards themselves or use AI-created dashboards regularly.Source: §3.1.3 · pp. 4–5

3.2 Design Goals

  • The design goals prioritize reuse of existing views, flagging underused content for user review, adapting prominence through interaction while retaining access, and preserving dashboard continuity.These goals respond to substantial creation and maintenance effort, limited ongoing maintenance, and the possibility that infrequently used content remains useful.Source: §3.2 · p. 5

4 Design Space: Living Dashboards

Living Dashboards are designed to evolve with users’ needs while preserving orientation and analytical context. Their design space specifies constraints and four dimensions that characterize how a dashboard observes use, adapts, times changes, and shares agency with users.

4 Design Space: Living Dashboards

  • A living dashboard’s views are born, wither, revive, and die in response to users’ needs and observed use; presentation changes can be autonomous, while consequential view-set changes require approval.Birth adds a view, withering reduces prominence, revival restores it as use resumes, and death is user-approved retirement from the active dashboard.Source: §4 · p. 5

4.1 Design Constraints

  • Design constraints keep the view set stable until user approval, retain access and legibility for underused views, and protect users’ mental maps through gradual or communicated changes.Adaptations affecting navigation or interpretation should be perceptible, including changes in a view’s location, size, encoding, or presence.Source: §4.1 · pp. 5–6

4.2 Design Dimensions

  • The design space is a four-dimensional tuple: Observation, Adaptation, Temporality, and Agency.Choices in one dimension can influence the others.Source: §4.2 · p. 6

4.2.1 4 Observation

  • Observation captures evidence of use and varies by interaction granularity, whether linked-view changes count as use, and whether activity is attributed individually or collectively.Fine-grained signals can provide richer evidence of intent but increase observation-model complexity; counting transitive interaction may inflate relevance, while ignoring it may undervalue supporting views.Source: §4.2.1 · pp. 6–7

4.2.2  Adaptation

  • Adaptation can change the view set with approval, emphasize views through size, alter encodings, rearrange layout, or vary visibility while retaining access and legibility.These five forms vary in intrusiveness according to the extent of change and the user’s task.Source: §4.2.2 · p. 7

4.2.3 Â Temporality

  • Temporality specifies when changes occur and how old activity loses influence: adaptation may be continuous, session-boundary, or threshold-triggered, while decay sets the dashboard’s temporal memory.Continuous adaptation may disrupt ongoing analysis; session-boundary adaptation avoids in-session changes but still requires users to understand changes when they return.Source: §4.2.3 · p. 7

4.2.4 § Agency

  • Agency specifies user understanding and influence: presentation changes may occur autonomously, additions and removals require approval, transparency explains changes, and reversibility enables overrides or undo.The design goal of preserving analytical continuity argues for at least partial reversibility.Source: §4.2.4 · p. 8

5 Implementation

The prototype adapts dashboard views using interaction and query signals, while reserving changes to the view set for user approval. It explains and supports reversing changes.

5 Implementation

  • The web-based prototype instantiates the paper’s design goals and constraints through a sidebar, focus-score and layout engines, recommendations, and a dashboard workspace.Its recommendations include previews; Figure 2 shows the workspace, sidebar, and header.Source: §5 · p. 8, Figure 2 · p. 10

5.1 Reusing and Adding Views

  • Natural-language queries first seek suitable existing views; identified views are highlighted and receive focus-score boosts, while missing views can be recommended for addition.Recommendations prioritize reuse, and additions require user approval.Source: §5.1 · p. 8

5.2 Estimating Use Over Time

  • Per-view focus scores combine clicks, pointer movement, dwell, and query boosts; weighted inputs and a decay half-life let scores reflect changing use over time.Lower-scoring views receive larger positive updates when re-engaged, while older activity loses influence.Source: §5.2 · p. 9

5.3 Withering & Reviving Views

  • The Layout Engine withers low-scoring views by shrinking and deemphasizing them, then restores size and visibility after renewed interaction without changing their grid position.The prototype chose a vignette treatment; hovering restores full opacity, and shrinking does not redistribute the vacated space.Source: §5.3 · pp. 9–10, Figure 3 · p. 10

5.4 Managing Changes to the View Set

  • Adding or removing a view requires approval: the Recommendation Engine proposes additions and flags views below the removal threshold, leaving the decision to the user.Removal recommendations include a reason and preview; a view dies only when the user approves its removal.Source: §5.4 · pp. 10–11

5.5 Explaining and Reversing Changes

  • Recommendations explain why a change is proposed, and a history lets users revert applied changes, keeping consequential adaptations understandable and recoverable.These mechanisms support transparency, reversibility, and a mixed-initiative interaction in which users retain control.Source: §5.5 · p. 11

6 Exploratory User Study

The exploratory between-subjects study compared Living Dashboard with an AI-guided baseline and examined task performance, workload, usability, and responses to adaptive behavior. Twelve university-affiliated participants were randomly assigned across the two conditions.

6 Exploratory User Study

  • The study compared the prototype with an AI-guided baseline lacking autonomous withering and revival, focusing on short-term use and perceptions of the behaviors encountered.Both conditions used the same underlying AI model, but Living Dashboard also supported AI-assisted view creation and highlighting; its decay half-life was accelerated to 45 seconds.Source: §6 · p. 11
  • The study asked how the conditions differed in analytical task performance, perceived usability and workload, and responses to autonomous lifecycle behavior.The research questions covered task accuracy, perceived usability and workload, and users’ perceptions of adaptive behavior.Source: §6 · p. 11

6.1 Participants

  • Twelve participants were randomly assigned equally to the conditions, with three from each of two regions in each group; experience ranged from novice to expert.Participants came from university settings and reported experience with visualization, dashboards, chat interfaces, and dashboard authorship before the study.Source: §6.1 · pp. 11–12, Figure 4 · p. 12

6.2 Dataset and Tasks

Participants used a modified Power BI revenue dataset to complete ten analytical tasks in individual online sessions. The study recorded task outcomes and collected questionnaire and qualitative responses, but did not use completion time to assess efficiency.

6.2 Dataset and Tasks

  • Ten tasks on a modified Power BI revenue dataset covered existing-view exploration and new-view creation, from information retrieval to comparative analysis requiring cross-filtering.Eight tasks required existing views and two required creating views; participants could complete tasks in any order or skip them.Source: §6.2 · pp. 12–13

6.3 Procedure

  • Participants completed the tasks individually online using think-aloud, followed by an interview and questionnaire; sessions lasted about one hour.They received an interface demonstration and familiarization time, and researchers recorded screen, video, and audio; session state was not carried between participants.Source: §6.3 · p. 13

6.4 Measures and Analysis

  • Analysis reported correct, incorrect, and skipped outcomes, participant-level task counts, questionnaire ratings, and qualitative observations reviewed manually by two authors.Task completion time was excluded from efficiency conclusions because it varied with exploration, AI use, and skipped tasks; qualitative review used no formal coding procedure.Source: §6.4 · p. 13

6.5 Results

In this exploratory sample, Living Dashboard participants answered more tasks correctly and reported lower workload and higher usability than baseline participants. Qualitative findings describe benefits and frustrations with view creation and adaptive behavior, while the study also found mixed reactions to shrinking views.

6.5.1 Qualitative Results

  • Living Dashboard participants used AI to create targeted views, while baseline participants often relied on familiar chart types and encountered difficulties on custom-view tasks.Participants valued the speed and preconfigured filters of AI-assisted view creation.Source: §6.5.1 · p. 14
  • Withering and revival helped some participants focus on single views, but made cross-view comparisons and revisiting views more cumbersome for others.Participants described revival as zooming in and out; both L5 and L6 resorted to browser zoom on comparison charts.Source: §6.5.1 · p. 14
  • Participants struggled with cross-filtering, and T6 had among the fewest correct answers; some overlooked or found the suggested Gantt chart complex.The authors noted that more structured onboarding for filtering techniques may be beneficial.Source: §6.5.1 · p. 14

6.5.2 Descriptive Quantitative Results

  • 9.17 versus 7.17 tasks were answered correctly on average in Living Dashboard and Baseline, respectively; overall, accuracy was 91.7% versus 71.7%.The condition totals were 55 of 60 assigned task instances correct for Living Dashboard and 43 of 60 for Baseline.Source: §6.5.2 · p. 14, Figure 5 · p. 15
  • Mental demand, effort, and annoyance ratings were lower in Living Dashboard, while perceived success was higher than in Baseline.Mental demand was 3.50 versus 5.00, effort 3.83 versus 4.50, annoyance 3.00 versus 4.67, and perceived success 4.50 versus 3.50.Source: §6.5.2 · pp. 14–15
  • Mean SUS scores were 76.25 versus 51.67, and satisfaction was 5.50 versus 4.33 for Living Dashboard and Baseline; participants’ ratings of shrinking were mixed.Five of six Living Dashboard participants rated it better than a static dashboard; three agreed shrinking suited their workflow, and four said it helped them focus.Source: §6.5.2 · p. 15

6.6 Scope of the Findings

The study offers an initial account of how participants used the prototype and responded to changing view prominence. Its combined features and study setting constrain what the comparison can establish.

6.6 Scope of the Findings

  • The study provides an initial account of prototype use and responses to changing view prominence during analytical tasks.The Living Dashboard condition combined AI-assisted view creation, highlighting existing views, and autonomous withering and revival.Source: §6.6 · p. 15
  • Figure 6 presents four seven-point task-experience ratings and individual SUS scores on a 0–100 scale, with condition means shown for SUS.Error bars on the task-experience ratings show ±1 SD.Source: Figure 6 · p. 16
  • The comparison cannot identify which combined feature accounts for observed differences in task performance and ratings.There were six participants per condition, one dashboard, brief sessions, and no observed retirement recommendation; the study therefore does not assess the Death stage or longer-term maintenance.Source: §6.6 · p. 16

7 Discussion

The discussion treats the quantitative comparison as an early signal, while emphasizing qualitative findings about AI use and changing view prominence. It identifies design tensions and limits that should shape future evaluation.

7 Discussion

  • The exploratory study found more correct tasks, higher usability, and lower workload with Living Dashboard, but the comparison does not isolate adaptive behavior.The authors regard the results as an early signal because the conditions differed in AI assistance as well as adaptive behavior and the study was small and short.Source: §7 · p. 16

7.1 AI as an Onboarding Assistant and Validator

  • Participants used AI not only for analysis but also to learn what the dashboard displayed and to validate their own reasoning.These observations come from the Living Dashboard condition and align with Dhanoa et al..Source: §7.1 · p. 16

7.2 Design Alternatives for the Withering Effect

  • Withering drew mixed responses: it helped focused single-view work but frustrated comparison, where participants wanted direct control and had to scroll between tiles.Suggested alternatives include layout reflow and explicit revival controls; accelerated decay may have amplified comparison-task frustration relative to slower real-world decay.Source: §7.2 · pp. 16–17

7.3 Interaction as a Signal for View Prominence

  • Mouse-interaction signals sometimes withered views that participants were still reading during comparison tasks.More observation, such as dwell, pointer proximity, or gaze, could capture missed attention but would make the model more complex and less predictable; explicit correction is another design opportunity.Source: §7.3 · p. 17

7.4 Use of Living Dashboard in Collaborative Settings

  • The evaluation studied individual use, although the system was designed for collaborative settings; voice input could use meeting discussion to generate or revive views.That collaborative scenario remains for future evaluation.Source: §7.4 · p. 17

7.5 Limitations and Future Work

  • Six participants per condition and a single dashboard limit generalizability; authors also had to configure decay and interaction parameters directly in code.The authors call for a larger, more controlled study and evaluation across diverse real-world dashboards and datasets.Source: §7.5 · pp. 17–18
  • Accelerated decay may have amplified comparison-task frustration, while the short study did not observe the Death stage or retirement recommendations.Longitudinal studies are needed to examine responses to slower decay and retirement recommendations over extended use.Source: §7.5 · p. 18
  • AI hallucinations affected task performance in a few cases in both conditions, which used the same underlying AI.The authors suggest dataset- or visualization-level constraints as a possible way to reduce dashboard-specific hallucinations.Source: §7.5 · p. 18

8 Conclusion

Living Dashboards adapt view prominence and lifecycle to observed use, with autonomous changes for lower-stakes actions and user decisions for consequential ones. In an exploratory study, participants answered more tasks correctly and reported lower workload and higher usability, while qualitative responses varied by task context.

8 Conclusion

  • Living Dashboards make views born, wither, revive, and die in response to how they are used, unlike dashboards whose structure remains fixed by design.The concept is formalized through Observation, Adaptation, Temporality, and Agency.Source: §8 · p. 18
  • The web-based prototype uses interaction-derived focus scores to manage view lifecycles, acting autonomously on withering and revival while leaving additions or retirements to users.Participants also responded positively to visual communication of view state and AI-assisted view creation.Source: §8 · p. 18
  • In the exploratory N = 12 study, Living Dashboard participants answered more tasks correctly and reported lower workload and higher usability, but conditions differed beyond adaptive behavior.Withering was helpful for focused single-view work and frustrating during comparison.Source: §8 · p. 18

Use of Generative AI

The authors used generative AI tools to support implementation coding and preparation of the visualizations in Figure 6. They reviewed and verified all outputs.

Use of Generative AI

  • Codex by OpenAI supported substantial implementation coding and the Python script for Figure 6 visualizations; authors reviewed and verified all outputs.Source: Use of Generative AI · p. 18
Loading 2610.08393v1…