Source-linked AI summary
Composition and Configuration Patterns in Multiple-View Visualizations
Xi Chen, Wei Zeng, Yanna Lin, Hayder Mahdi Al-maneea, Jonathan Roberts, Remco Chang
TL;DR
Multiple-view visualizations are widely used, but the design space has few specific guidelines. This paper analyzes 360 annotated MV designs to identify composition and configuration patterns and build an interactive recommendation system; a user study suggests the system helps novices and experts understand the design space and promote appropriate designs.
Problem
Despite the widespread use of MVs, there are few guidelines for their design, and complex layouts can require tedious manual curation.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
The findings support interactive exploration of existing MV designs and recommendations for designing MVs.
Source: §1 · p. 2, §8 · p. 9The 14-type view taxonomy is not exhaustive, and some compound views must be categorized by their main view type.
Source: §7 · p. 9
Abstract
from arXiv · showhide
Multiple-view visualization (MV) is a layout design technique often employed to help users see a large number of data attributes and values in a single cohesive representation. Because of its generalizability, the MV design has been widely adopted by the visualization community to help users examine and interact with large, complex, and high-dimensional data. However, although ubiquitous, there has been little work to categorize and analyze MVs in order to better understand its design space. As a result, there has been little to no guideline in how to use the MV design effectively. In this paper, we present an in-depth study of how MVs are designed in practice. We focus on two fundamental measures of multiple-view patterns: composition, which quantifies what view types and how many are there; and configuration, which characterizes spatial arrangement of view layouts in the display space. We build a new dataset containing 360 images of MVs collected from IEEE VIS, EuroVis, and PacificVis publications 2011 to 2019, and make fine-grained annotations of view types and layouts for these visualization images. From this data we conduct composition and configuration analyses using quantitative metrics of term frequency and layout topology. We identify common practices around MVs, including relationship of view types, popular view layouts, and correlation between view types and layouts. We combine the findings into a MV recommendation system, providing interactive tools to explore the design space, and support example-based design.
1 INTRODUCTION
Multiple-view visualizations are widely used, but designers have few specific guidelines for arranging them. This paper studies real-world MV composition and configuration and uses the findings to build an interactive recommendation system.
1 INTRODUCTION
- Designers have few specific guidelines for MVs, and complex layouts often require tedious manual curation through trial and error.Predetermined templates serve common data types, but complex data and tasks can still produce layouts that fail to meet guidelines.Source: §1 · p. 1
- The study characterizes composition by view-type frequency, diversity, and correlation, and configuration by view position and size.The paper codes each view's type and bounding box, then describes the overall MV using these two pattern dimensions.Source: §1 · p. 1
- The study analyzes 360 annotated MV designs from IEEE VIS, EuroVis, and PacificVis publications from 2011–2019.The corpus is used for quantitative analysis of composition and configuration patterns.Source: §1 · pp. 1–2
- Analysis identifies composition and configuration patterns, including view-type frequencies, aspect ratios, and positions.Most view-type aspect ratios fall within [1/2, 2], except for types such as Area and Panel.Source: §1 · p. 1
- The interactive recommendation system supports exploration and design recommendations, and a user study suggests it helps novices and experts understand the design space and promote appropriate MVs.The study presents the system as a way to explore existing MV designs and recommend designs from view-type and layout information.Source: §1 · p. 2
2 RELATED WORK
Prior MV research has developed guidelines, tools, and ways to describe relationships between views, while data-driven design work has largely focused on single visualizations. This paper frames its empirical study as a foundation for understanding and designing multi-view layouts.
2 RELATED WORK
- MVs present different perspectives of data, but existing guidance and layout curation leave designers with a difficult, laborious task.Prior work includes general guidelines, authoring tools, and formal categories for arrangements such as juxtaposition, superimposition, overloading, nesting, and integration.Source: §2 · p. 2
- The paper studies real-world MV patterns because human-created layouts are not arbitrary, aiming to establish a data-driven foundation for MV design.The authors characterize these patterns as composition and configuration.Source: §2 · p. 2
- Data-driven visualization methods largely map data and task inputs to a single visualization, while this paper targets composition and configuration across multiple views.Prior examples include learning visualization rankings, generating visualizations, and learning design constraints.Source: §2 · p. 2
- The analysis borrows treemap metrics—aspect ratio, stability, and relative-position-change—to examine MV configuration.The paper motivates this choice by the shared concern with efficient use of space.Source: §2 · p. 2
3 DEFINITION AND VIEW SPECIFICATION
The paper models an MV as a display-space layout of two or more views, each specified by view type and bounding box. It also represents small multiples hierarchically, with the corpus reaching at most two view levels.
3 DEFINITION AND VIEW SPECIFICATION
- An MV is a layout arranging two or more views in a display space, with each view represented by its type and bounding box.The bounding box specifies normalized center position and size, while the arrangement of boxes defines the configuration pattern.Source: §3 · pp. 2–3
- The study uses 14 view types: 12 information-visualization types, plus SciVis and Panel.SciVis groups scientific visualizations; Panel covers graphic widgets such as menus, legends, and narrative text not overlaid on other views.Source: §3 · p. 3
- Small multiples form nested view groups, and the corpus's view hierarchies reach at most level two.The paper represents an MV as views or groups of views, with each view specified by its type and bounding box.Source: §3 · p. 3
4 DATASET CONSTRUCTION
The dataset is assembled by collecting MV images from visualization publications, manually annotating their views, and refining bounding boxes into layouts without gaps or overlaps.
4.1 Multiple-View Visualization Collection
- The authors extract 16,891 figures from 1,976 papers, then manually select eligible MV images, yielding 360 images from 2011–2019.The corpus covers IEEE VIS, EuroVis, and PacificVis; IEEE VAST contributes 197 of the 360 MVs.Source: §4.1 · p. 3, Table 1 · p. 3
4.2 View Annotation
- Because automated view detection produced unsatisfactory results, the authors manually annotate view types and bounding boxes using a dedicated tool.The tool supports drawing and labeling views, editing their attributes, and reviewing the dataset; annotations retain normalized position and size.Source: §4.2 · pp. 3–4
4.3 Layout Refinement
- A bottom-up refinement groups small multiples and neighboring boxes, aligns or resizes them, and repeats until the layout has no gaps or overlaps.The alignment process uses a threshold of 3% of the group box's average width and height.Source: §4.3 · p. 4, Figure 3 · p. 4
5 COMPOSITION AND CONFIGURATION ANALYSIS
Across 360 multiple-view designs, the analysis describes which view types co-occur, how layouts vary by venue, and where view types tend to appear. It finds recurring patterns in view counts, layout complexity, aspect ratios, and relative positions.
5 COMPOSITION AND CONFIGURATION ANALYSIS
- 62.2% of the 360 designs contain no more than five views; three-view designs are most common, and Panels appear in 68.3%.Bar, Network, and Line views each appear in more than 30% of designs, while Circle appears in 1.6%.Source: §5 · p. 4
5.1 Composition Pattern Analysis
- Panel appears with each of the other 13 view types at conditional probabilities above 0.5, while Bar appears with SciVis in only 0.07 of cases.P(Panel|SciVis) is 0.8, the highest conditional probability reported; co-occurrence probabilities can also be asymmetric, as P(Bar|Circle) is 0.5 and P(Circle|Bar) is 0.03.Source: §5.1 · p. 5, Figure 5 · p. 5
5.2 Configuration Pattern Analysis
- 98 layouts emerged; the top ten use four or fewer views and cover 85.7% of InfoVis designs versus 40% of PacificVis designs.The top ten account for 63.8% in EuroVis, 55.6% in SciVis, and 44.2% in VAST; 2A is more prevalent in InfoVis and SciVis than in the other venues.Source: §5.2 · p. 5, Table 2 · p. 5
5.3.1 View Type & Aspect Ratio
- Aspect ratios for most view types fall within [1/2, 2], with Area charts usually above 1 and Panels usually below 1.The authors report horizontally long, narrow Area charts in small multiples and vertically long, narrow Panels at the sides of displays.Source: §5.3.1 · p. 6, Figure 7 · p. 6
- Bar and Network view aspect ratios average near 1, Distribution ratios peak around 1/2, and Panel ratios peak around 1/3.Bar charts also include narrow examples near 1/5 and 5; Network ratios are concentrated within [1/2, 2].Source: §5.3.1 · p. 6
5.3.2 View Type & Relative Position
- Diag., SciVis, and Network views occupy over 25% of display space on average, while Area and Bar occupy about 5%; Circle occupies about 25%.The authors suggest Area and Bar may receive small spaces because they typically depict summary statistics; Circle’s larger average reflects infographic designs, while it is smaller in other visual analytics MVs.Source: §5.3.2 · p. 6
- Panels rarely occupy the center, whereas Diag. views are mostly central; large views tend toward the top-left and center, and most reported stability values are below 0.2.Panel positions and sizes are rather stable across the top ten layouts, while Text in 4H, Point in 4C, and Map in 2B have stability values above 0.3; Map is more stable in 2A than 2B.Source: §5.3.2 · p. 6
6 RECOMMENDATION SYSTEM
The recommendation system turns composition and configuration findings into interactive tools for exploring existing multiple-view designs and recommending layouts from user input. Two user studies evaluated the modes, with exploration supporting accurate pattern queries and full design mode rated easier to use despite no significant completion-time difference among modes.
6 RECOMMENDATION SYSTEM
- The system offers two modes because the analyses are difficult to express as clean guidelines: one explores existing designs, while the other recommends layouts from user input.The design mode searches the 360-design database for layouts matching the user's drawn design.Source: §6 · p. 6, §6.2 · p. 7
6.1 Exploration Mode
- Exploration mode filters existing designs by view type, number of views, and layout, then presents results as grouped or colored dots with publication details on hover.Users can union selections within a facet and intersect selections across facets.Source: §6.1 · pp. 6–7, Figure 9 · p. 7
6.2 Design Mode
- Design mode lets users add, remove, resize, position, and align views, then ranks database layouts by mutual information with the user's 3×3-grid representation.The system represents view-type proportions across grids as 3×3×14 tensors; finer grids could improve proximity but increase computation, so it uses 3×3.Source: §6.2 · pp. 7–8
6.3 User Study
- The user evaluation comprised two virtual studies, one assessing exploration and one assessing design mode.The design study involved 13 visualization practitioners with at least four years of design experience.Source: §6.3 · p. 8, §6.3.2 · p. 8
6.3.1 Study 1: Qualitative Evaluation for Exploration Mode
- In the exploration study, all 20 participants completed 10 pattern questions in eight minutes, with 97.8% average accuracy and positive ratings for usefulness and ease of use.Participants suggested reducing overused colors and adding facets such as authors and research topics.Source: §6.3.1 · p. 8
6.3.2 Study 2: Quantitative Evaluation for Design Mode
- Full mode was rated easier to use (mean 6.6) than Partial (5.2) and Basic (3.5), but design mode had no significant effect on completion time.Task complexity did affect time: arranging 3 views was faster than 5, which was faster than 7.Source: §6.3.2 · p. 9, Figure 12 · p. 9
7 DISCUSSION, LIMITATION, AND FUTURE WORK
The analyses identify recurring composition and configuration patterns in multiple-view visualizations, while the authors note limits in view-type coverage, design factors analyzed, and dataset scope. They describe future work on richer view representations, semantic structure, consistency, and design goals.
7 DISCUSSION, LIMITATION, AND FUTURE WORK
- Most MVs contain fewer than five views; Bar and Line charts are common, layouts are usually simple, and view types show typical aspect ratios and positions.VAST papers tend toward more complex layouts than InfoVis and SciVis papers; Panels are typically not centered, while Diagrams tend to be.Source: §7 · p. 9
- The observed composition and configuration patterns provide a foundation for a recommendation system intended to assist designers in choosing appropriate MV designs.Usability and effectiveness were supported by feedback from visualization novices and experts.Source: §7 · p. 9
- The analysis categorizes views into 14 types, but this is not exhaustive and can classify compound views only by their main type.The study analyzes view types and layouts, not underlying selection mechanisms such as data, tasks, preferences, or perceptions.Source: §7 · p. 9
- The recommendation system draws only on a small corpus from visualization conferences, which includes few designs for large displays.The authors propose expanding the corpus with designs from other venues, the Internet, and authoring tools.Source: §7 · p. 9
- Future work includes analyzing semantic links between views, evaluating visual consistency, and connecting dashboard patterns to explicit design goals.The authors also propose using real-data views instead of view-type icons.Source: §7 · p. 9
8 CONCLUSION
The paper combines an empirical study of multiple-view designs with an interactive recommendation system based on the analysis. Its 360-image dataset and fine-grained annotations support both exploration of existing designs and recommendations based on patterns and user inputs.
8 CONCLUSION
- The study analyzes view-type associations and layout arrangements using fine-grained annotations of 360 MV images from IEEE VIS, EuroVis, and PacificVis publications from 2011 to 2019.The authors plan to release the dataset for future MV-design research and other recommendation systems.Source: §8 · p. 9