Essay

Essay on Visualization Perception

Tables, charts and various types of graphs are used in textbooks and we are familiar with them since we started school. Not only do we read them and gather information conveyed through them, but also do we construct graphs ourselves, both manually using pencils and rulers and using computer tools such as R, Excel and Tableau. As an analyst who uses charts on daily basis and includes them as a part of our products delivered to clients, also a student majoring in Analytics, I use data visualization in both my work and school projects and found them very powerful and efficient at getting through the message in a shorter time. Moreover, data visualization is an essential part of presentations today in a wide array of industries. Nonetheless, it does not promise that all data visualization would aid with presenting the data accurately and efficiently, unless we understand the role of perception in visualization and incorporate that into selecting and constructing visualization that suits our needs.

The purpose of using visualization in data presentation is to help to clearly display the data and convey the message, for example the relationship/comparison between variables, to readers and they could absorb the information quickly. The process of readers to analyze and organize the information and interpret its meaning leads to readers’ perception, and thus, whether the perception really aligns with presenters’ intention is very critical to communicating the accurate message and ensuring readers could successfully receive it.

Is that always the case that readers could understand the data without difficulties via various types of visualization? Unfortunately, no. Cleveland and McGill performed experiments using elementary perceptual tasks, there would always be errors, and they found that tasks involving different judgements have different levels of accuracy (1984). Position judgements are more accurate then length and angle judgement while length is more accurate than angle judgement. Herr and Bostock, in their study on crowdsourcing using Amazon’s Mechanical Turk further confirmed the experiment results from Cleveland and McGill. They found similar distribution of errors using a different population of subjects after trying the replicate past experiments.

Knowing that not all visual aids help to present data and information for readers to process it accurately, perception makes it important for us to select the most effective way to present our data so that there is no confusion or misunderstanding and our visual aids are really helping the presentation/research report. What would be the appropriate approaches when selecting visual aids?

Cleveland and McGill, in their study, also provide insights to this. Besides that position judgement tops all in terms of accuracy, many new – at the time of the research – visualization methods were proposed by Cleveland and McGill and they were proven to lead to more accurate perception. Dot charts and bar charts could replace divided bar chats and pie charts. When the angle differences are small in the pie chart or the length differences are small in the bar chart, it would be hard for readers to tell the relationship amongst the groups, with a dot chart, however, it is very easy to tell order of ranking. For statistical maps where shading is used to represent the frequency in different regions on the map, frame-rectangle charts could do a better job. The shading might be misleading due to the unequal areas of different regions. Additionally, instead of plotting curves to show differences, plotting the difference directly would show the trend more clearly because it is challenging for readers to process the curves and calculate the area over time between the two. Similarly, Few, in his article published on Perceptual Edge, suggested selecting the right chart type is critical. We could compare below two charts and clearly tell the bar charts on the right tell a better story then the charts on the left (Few, 2004). From the Cost vs. Revenue chart on the left, it is hard to tell the magnitude of differences between costs and revenue, for January, when Revenue is smaller, we cannot even see the green dot. On the 2004 YTD Expenses line chart, it is hard to compare across departments and rank them in any order. Line charts work best for trends across time, not relationship between values.

Figure 2: Poor graph choice on the left vs. an effective choice on the right.

Moreover, as we move into the century of internet, a lot of experiments would be done through computers, Herr and Bostock’s study provides some guidelines when conducting experiments and collecting data online. Comparisons of rectangles with aspect ratios of 1 would result in highest error than other ratio combinations, we should avoid using squares as a result. Also, gridlines should be spaced more than 8 pixels apart and the ideal chart height is at 80 pixels, anything above that does not improve perception accuracy.

Hence, we learnt from Cleveland and McGill that several new charts, such as dot charts, could be adopted because they result in higher accuracy as compared to traditional charts. We also learnt that when presenting data on monitors, we could follow the setting suggested by Herr and Bostock so that it lead to highest possible accuracy.

Do we always need charts/graphs? Not necessarily. Tables are also useful for data presentation (Few, 2004). When the shape of data contains messages to be conveyed or we need to show the relationships of values, graphs would be helpful. If we were to show or compare specific values and need to be precise, tables are more appropriate (Few, 2014). Thus, thinking ahead the purpose of data presentation is important.

Furthermore, after we are able to land at an appropriate graph for data visualization, there are other techniques we could utilize. First, we need to keep in mind that readers have limited short-term memory (Few, 2004). Few published another study on Information Week, and elaborated how to maximize the power of data visualization. Short-term memory is where the work of sense-making is processed. Readers’ ability of processing the data provided in the graphs is limited to the size of short-term memory. Figure 3 is an example of a graph that contains more information of what could be process in short-term memory. Instead of including all info in one chart, it might work better to break into a few so that readers could process the information.

Figure 3: Example of a graph that exceeds the limits for short-term memory.

In his work published on Information Week, Few also explored another technique: “preattentive” processing. Few stated: “in contrast to the conscious part of perception, which is called ‘attentive’ processing.” Preattentive processing is “extremely fast and broadband in that we can simultaneously perceive a large number of these basic visual attributes called ‘preattentive attributes’.” Because preatentive perception is done in tandem, it is faster than attentive processing. Figure 4 below shows attentive processing, and Figure 5 shows an example of preattentive processing. If readers need to count the number of “5” in the chart, it takes much longer to count in Figure 4 as compared to Figure 5.

Figure 4: Example of attentive processing.

Figure 5: Example of preattentive processing.

Figure 6: Preattentive attributes of visual perception most applicable to data presentation.

More examples of preattentive attributes are shown in Figure 6 above, with the aid of the differentiation, readers could process information much quicker. We need to be careful with preattentive attributes though, Figure 7 is an example of a misuse of different colors and it makes it confusing and not easy for readers to process the information.

Figure 7: Example of a misuse of hue for the display of quantitative values. (Notes: This is a screen capture of a graph that was constructed using interactive examples on Visualize’s Web site.

In conclusion, perception affects how readers receive and process data, and in turn affects how visualization should be done. Understanding how perception changes to different data presentation is critical for us to choose appropriate visual aids to present data effectively and let readers understand accurately. We could replace some of the tradition charts by new one such as dot charts, as suggested by Cleveland and McGill’s experiment. Also, appropriate charts should be selected to present the data. When using computers to present data, Herr and Bostock’s study suggest pixels to be used on the screen. Additionally, we should limit the amount of the information contained in graphs and use preattentive attributes in a constructive way for visualization enhancement, as suggested by Hew. As new technology keeps advancing, more visual aid tools will be adopted in the future and help us to translate the information, AR is a great example. There might not be a perfect way to achieve complete accuracy, but getting more and more accurate is a significant and meaning step forward.

References

Stephen Few; Information Week, Data Presentation: Tapping the Power of Visual Perception, (Aug., 2004)

Stephen Few; Perceptual Edge, Common Mistakes in Data Presentation, (Sep., 2004)

Jeffrey Herr; Michael Bostock, Crowdsourcing Graphical Perception: Using Mechanical Turk to Assess Visualization Design, Jeffrey Heer and Michael Bostock

William S. Cleveland; Robert McGill, Journal of the American Statistical Association, Vol. 79, No. 387. (Sep., 1984), pp. 531-554.