B. Bach, E. Freeman, A. Abdul-Rahman, C. Turkay, S. Khan, Y. Fan, and M. Chen. Dashboard design patterns. IEEE Transactions on Visualization and Computer Graphics, 29(1):342-352, 2023. DOI. (Presented at IEEE VIS 2022, and IEEE VIS 2022 Honorable Mention.)
This paper was led by Benjamin Bach and included contributions from several members of the Generic Support Team of the RAMPVIS project. It was partly inspired by the significant role played by dashboard-based visualization tools during the COVID-19 pandemic around the world. During the early stage of the RAMPVIS project, there was debate within the team as to the apparent conflicts between some common design features (e.g., large numbers as details first) and common VIS guidelines (e.g., overview first). I gained my appreciation of showing some big numbers in dashboards from two perspectives: (i) An industrial application in the 2000s, where the dashboard-based visualization provided by two Swansea VIS researchers (Dr. Simon Walton and Dr. Shoukat Islam) was hugely appreciated by the users; (ii) my information-theoretic analysis of Shneiderman's Information-Seeking Mantra "overview first, zoom, details on demand", which led to a better understanding of the merits of some details-first solutions (i.e., when users has already had the knowledge gained from overviews during the previous days, weeks, or months). When Ben asked me to write something for the paper, I noticed that the categorization scheme being constructed by Ben and the colleagues could help abstract reasoning about such merits. I drafted the texts in Section 7.1 and drew Figure 10 for the original submission. Later, the analysis of the big numbers had to be removed largely due to the page limit. Here we reuse the texts and the relevant figures to demonstrate the IVAS approach.
From an information-theoretic perspective, a dashboard encodes a data space that is smaller than the data space of the data to be displayed, but larger than the data space of the decisions/tasks performed by users when they view the dashboard. Hence, information loss in the data space of the dashboard is inevitable. Such information loss may be positively referred to as abstraction and filtering; these facilitate effective abstraction close to the data space of the decisions/tasks and efficient cost reduction in interactive visualization. However, information loss may also negatively cause confusion, misinterpretation, or erroneous judgment.
Figure 5: Example of using tabs in dashboards. Clicking on of the tabs below the main view leads to a similar view with a different data set (e.g., deaths (left) and cases (right)).
Figure 10: An example of exploring design patterns using a systematic optimization process [14].
For example, consider a time series representing the daily number of positive test cases for Covid-19 during a period of 500 days. Figure 5 shows a design with four visual encodings (number, trend-arrow , signature chart and visualization ) for the same time series data. They all lose information, but in different ways: the line chart visualization may lose information through its limited height and the vertical pixel-resolution limits the range of data visible without scrolling. The large number shows the latest figure (a single-value ), while omitting all other data points in the time series dataset . The trend error (a derived value) and signature charts are different levels of abstraction between the number and full visualization, representing the full range of abstraction in one dashboard view.
Few [21] suggests showing numerical values as numbers to convey the data value accurately and quickly, as compared to, e.g., trying to perceive the last data point in the line chart or using a tooltip. Meanwhile, while the line chart shows the long and medium-term trends well, perceiving short-term trends can be difficult. The arrow next to the large number addresses this difficulty in terms of speedy observation, though it is not as precise as showing a number that represents the penultimate data value, or the delta between the last two data points.
Interestingly, numbers and arrows may seem redundant, especially if the line chart offers tooltip interaction. The primary reason for including both must be the reduction of cost in terms of both time and cognitive load in observing key indicators. For the latest number, the precision of this key indicator must be important to the users’ decisions/tasks. A dashboard showing only a line chart for the time series will likely frustrate users by requiring interaction to find key indicators, adding the cost of using, e.g., tooltip interaction. The interaction cost is not only due to the mouse movement to activate the tooltip but also due to the cognitive load for remembering the value after the tooltip disappears.
It is clear from this example that there are costs and benefits associated with different levels of abstraction and their visual encodings, perhaps motivating dashboards to combine several. However, the tradeoff here is the excessive cost of screen space when displaying several redundant visual encodings, and the increased cognitive cost of interpreting different levels of abstraction over the same data. It is necessary to display fewer numbers and visual information, to reduce the visual complexity of the dashboard. Designers thus need to find an optimal balance between abstraction and cost, e.g., to encode the latest data value(s) as numbers, and rely on perceptually less-precise plots to present overviews while reminding users of their past observations. As shown in Figure 10, with careful reasoning, exploring our dashboard design space does not necessarily involve exclusively considering many design options in a combinatory manner; rather, a considered combination can guide dashboard design.
[14] M. Chen and D. S. Ebert. 2019. An ontological framework for supporting the design and evaluation of visual analytics systems. Computer Graphics Forum, 38, 3, 131–144.
[21] Stephen Few. 2006. Information Dashboard Design: The Effective Visual Communication of Data. Vol. 2. O’Reilly, Sebastopol, CA.
Let us start with the Visualization block at the top of Figure 10 by considering a basic dashboard showing a time series plot depicting some COVID-19 data (e.g., number of daily cases). In many occasions during the RAMPVIS project, I asked colleagues who were working on dashboards to imagine a scenario, where a regional healthcare manager woke up in the morning, and wanted to see the latest key indicators relevant to the region while brushing her teeth.
Symptom: The "difficulty to see the actual values" can be translated to (i) the high cognitive cost to see as one has to relate each interesting data point on the time series with the y-axis to determine its value, and (ii) the high level of inaccuracy as visually mapping a visually encoded object back to a numerical value is usually not accurate. In abstraction, these two symptoms are Vis-High-Cost and Vis-High-PD respectively.
Cause: In terms of the number of data points (considering each data point as a variable), there are many data points, i.e., not enough alphabet compression (Vis-Low-AC). In terms of the number of distinguishable values for each data point, the perception of time series loses resolution, i.e., too much compression (Vis-High-AC).
Remedy #A: Tooltip is an interactive technique that can address the cause, that is, the location of the tooltip filters out other data points (Int-High-AC), and the pop-out small window for showing a numerical value reduces the compression of data resolution (Vis-Low-AC).
Side-Effect: Tooltip interaction can be costly and inconvenient (Int-High-Cost), especially during "tooth brushing" and the perceived value is easy to forget once the pointer device moved away from the interesting data point.
Cause: The action for the tooltip interaction is sensitive to small movement of the pointer device and thus demands high cognitive load (Int-High-Cost + Int-High-PD), the sensitivity is partly caused by the large number of locations available to be selected (Vis-Low-AC). The "easiness to forget" is partly due to the memory cost as humans have limited working memory, demanding more cognitive load to memorisation during interaction (Int-High-Cost + Int-High-PD).
Remedy #B: To display numbers directly without any interaction. The numerical values reduce the compression of data resolution (Vis-Low-AC).
Side-Effect: The number of numerical values that can be shown is small. When one needs to see many numerical values, one will encounter the difficulty to observe and reason about many numbers (Vis-Low-AC + Vis-High-PD) and/or to display many numbers within a limited display space (Vis-High-Cost).
Remedy #C: Only select a small subset of numerical values to be displayed (Vis-High-AC). Large fonts are often used to reduce the cognitive load (Vis-Low-Cost) The approach is in general only effective for users who use dashboards frequently and their need to see the latest data values is largely unchanging.
Remedy #D: Display many data points visually, e.g., using a time series plot (which is the basic dashboard at the beginning of the discussion). In terms of the number of data points, this is Vis-Low-AC, and in terms of the data resolution, this is Vis-High-AC.
Many dashboard designs feature a combination of Remedies #C and #D, or a combination of Remedies #C and #A.