Graphical Perception of Multiple Time Series

Editor's Notes

• #2: GraphicalPerception of Multiple Time Series Presented by Waqas Javed Bryan McDonnel And Niklas Elmqvist @ Purdue University
• #4: In this talk I will first discuss what we mean by the graphical perception Then will highlight the motivation for this work Next will talk about an overview of our work And some of the prior work in this field Then I will discuss commonly used visualization techniques for multiple time series Followed by the design of user study to measure the effectiveness of these techniques And last but not the least discussion about the results we obtained from the user study
• #6: Graphical perception is defined as The ability of users to comprehend the visual encoding and thereby decode the information presented in the graph. [click] When the graph have multiple time series, we talk about the graphical perception of multiple time series /* Graphical perception of multiple time series can be defined as The ability of users to comprehend the visual encoding and thereby decode the information presented in the graph, representing multiple time series. */
• #8: Motivation: [Click]Line graphs are today one of the most common types of statistical data graphics. They are used to visualize temporal data in a wide array of domains such as finance, politics, science, engineering, and medicine. // verbal bridge here [Click]Common tasks involving time series data often involve many concurrent series. Consider a stock analyst surveying the history of a set of stocks in an effort to find the next investment. This comparison will have to be conducted across each of the time series representing each individual stock. [Click]Graphical perception of multiple series plays an important role in the success of temporal visualizations used for such tasks. [Click]Effective guidelines are required for designers who need to find a suitable method when building a visualization application that support comparison across multiple time series.
• #10: We evaluate graphical perception of multiple time series as a function of different visualization types, under different conditions [Click] In particular, we evaluate the effect of following conditions [Click] Visualization type [Click] Number of series [Click] Available space [Click] Task type I will discuss each of these in detail later
• #12: Evaluation of graphical perception for statistical data graphics has a long history, originating from even before there were computers and graphics to turn charts into interactive visualizations. To our knowledge their exist no prior work that consider the graphical perception of multiple time series as a function of the parameters presented in the previous slide. [Click] Croxton et al. compared bar charts with circle diagrams and pie charts, and discussed the relative merits of these visualization techniques to perform the comparison tasks. [Click] Cleveland and McGill (1984) formalize the use of graphical perception for measuring the effectiveness of various graph techniques such as bar charts and pie charts [Click]Simkin and Hastie compared the accuracy of judgment while using simple bar charts, divided bar charts, and pie charts. Their findings are based on the comparison and estimation tasks, involving only two charts at a time. [Click]Lam et al. did investigate graphical perception of multiple line series, but their study focuses more on differences between low-resolution and high-resolution visual representations than on comparing the performance of line graph techniques. [Click]Heer et al. performed two controlled experiments to measure the effect of chart size and layering on user performance while performing discrimination and estimation tasks on data.
• #14: To classify the line graph visualization techniques we identify five different factors, suited for our experiment. These include [Click] Space management: that describes whether space is “shared” or “split” between time series. Shared space is typically more amenable to comparison between series (because they are overlaid in the same space), while data in split space may be easier to perceive (due to less clutter). [Click] Space per series: This factor defines the amount of vertical display space allocated to each individual time series. [Click] Identity: We often have to use graphical attributes such as color, or line style to convey identity. This factor is often more important for “line” techniques than for “area” techniques. [Click] Baseline: Comparison between time series is made easier with a “common” baseline than for an “individual” baseline, or one based on the “previous” time series displayed. [Click] Visual clutter: The clutter associated with the visualization technique plays an important role for large values of time series.
• #15: Now I will briefly discuss five different line graph time series visualization techniques based on the factors highlighted in the previous slide Simple line graphs Most commonly used line graph visualization where time is mapped to the horizontal-axis, and the value is mapped to the vertical-axis. Adding multiple time series is easy—just assign each series a unique graphical property, such as a color or a line style, and then add them to the shared space. As seen in Table, simple line graphs use a common baseline in shared space, making comparisons across series simple. However, because each series is represented by a line, distinguishing the series is challenging.
• #16: Small multiples applied to line graph visualization means that instead of adding all time series to the same graph space, we split the space into individual graphs, one for each time series. It is important that all charts use the same axis-scaling to allow easy comparison across the charts. Although such a technique decreases visual clutter but this allocates less vertical resolution to each individual time series.
• #17: A stacked graph is a shared space technique where one time series uses the value of the previous series as a baseline (the first series will use the origin of the graph as a baseline). Figure shows an example stacked graph for four time series. Because of the curious use of variable baselines, stacked graphs can use filled areas instead of lines to ease identification. space allocation for each graph is proportional to the sum of values of all time series.
• #18: Horizon graph. This visualization technique was originally presented by Saito et al. under the name “two-tone pseudo coloring”. [Click] Starting with a simple line graph, fill the area beneath the curve with a blue color for values above baseline, and a red color for values below baseline. Split the value range into discrete ranges, or bands, [Click] and mirror the negative values above the baseline. [Click] Final step introduces the notion of virtual resolution by wrapping the graph space using the bands, as discussed by Heer et al..
• #19: This virtual resolution and wrapping of negative values means that more space can be allocated for each individual time series despite the fact that horizon graphs use split space—instead of S/N, the space allocation for small multiples (other split space technique), horizon graphs achieve S/N times twice the number of bands. Just like small multiples, the split space layout means that the visual clutter is low.
• #20: The main reason why time series in simple line graphs can be difficult to identify is that the identifying graphical properties are restricted to a single (often thin) line. If we could somehow fill the whole area beneath the line, it will help the viewer distinguish between different time series. However, turning the lines into filled areas means that one curve might hide the other. [Click] We introduce Braided graphs to solve the problem by identifying the intersection points in time where two series change value ordering. [Click] Each filled area representing a series is cut into segments at these intersection points, and the individual segments are then depth-sorted and drawn with the highest value segment first.
• #21: Braided graph technique maintains common baseline. However, the resulting graph has a potentially high visual clutter for large numbers of series.
• #22: We performed a user study with 16 graduate students to measure the performance of these techniques.
• #23: In particular validating the following hypothesis [H1] Strength of the shared space techniques is that they permit easier direct comparison across series for a small visual span. Therefore, we predict that shared-space techniques (braided and simple graphs) will have better completion time for this kind of tasks. [H2] Our pilot study indicated that for larger visual spans, overlap and visual clutter will become a major factor for shared-space techniques. Split-space techniques, on the other hand, avoid occlusion, and we thus predict that they will have a better time performance for this kind of tasks. [H3]Many concurrent time series will cause decreased performance. [H4]Small display space will cause decreased performance. We also predict that the amount of vertical display allocated to each visualization will have a direct effect on user performance.
• #24: For the experiment we wanted to include tasks that are representative for common uses of temporal visualization. For this we included three different tasks in the experiment. [Click] Maximum task (local task) This task required the participants to find the time series with the highest value at a specific point in time. [Click] //Discrimination (global task) The discrimination task consisted of having the user determine which time series had the highest value at a point specific to each series. [Click] Slope (global task) Assessing the global slope requires users to find the time series with the highest increase during the whole displayed time period.
• #25: We designed the study as a within-subjects full factorial analysis on [Click] Visualization type: we included all the visualization types discussed except stacked graphs as our pilot study show that stacked graphs were mostly unsuitable for the tasks studied in this experiment. [Click] Tasks (T) [Click] Number of Series For the experiment we used 2, 4 and 8 time series [Click] Total chart size (based on pilot) we used three different chart sizes During the experiment we recorded time and correctness measures for each trial // For the experiment we used two repetitions per each condition generating 216 trials per participant and a total of 3,456 trials for the complete experiment
• #26: In the experiment we used two repetitions per each condition . For the analysis we used the average of the two repetitions while analyzing the measured data.
• #27: This figure shows the correctness measure for different tasks as a function of the visualization type. [Click] We found no significant effect of visualization type for correctness measure of the tasks, it indicates that the participants were equally careful, regardless of line graph type. However we did observe a significant effect of number of time series, total chart size and task type on the correctness measure, confirming H3 and 4. [Click] We can see the effect of task type in this figure as well. [Click] Participants performed best for the Max tasks, identifying them as the easiest one. While slope tasks turned out to be the most difficult ones.
• #28: // The time to complete a trial was measured from when the charts were first displayed to when the user clicked the Okay button on the answer dialog. We found that the time samples violated the normality assumptions of the analysis of variance, so we analyzed the logarithm of the times through repeated measures analysis of variance RM-ANOVA. // This figure shows the effect of number of time series on completion time for different task types. [Click] Our analysis confirmed that the number of time series has a significant effect on the completion time, as predicted in H3.
• #29: Effect of chart size on completion time for different task types, is shown in this figure.  [Click] We observed no significant effect of chart size on the completion time of different tasks. Partially invalidating H4 It appeared that since participants were asked to complete each task as quickly as possible and thus they tended to use the same amount of time regardless of chart size. But this behavior of their was manifested in decreased correctness, on which chart size did have a significant effect; In other words, a classic time/accuracy trade-off.
• #30: These graphs show completion times for each task as a function of the visualization type. [Click] We found that visualization type had a significant effect on the completion time for different tasks. [Click] We analyzed this further using a post hock Tukey HSD test; the figure at the bottom shows pair-wise relations for all tasks having a significant time difference. [Click] We observed for the discrimination tasks split space techniques performed significantly better than split space techniques. [Click] While for the Max tasks split space techniques outshined shared space techniques. [Click] There was a mix trend for the slope tasks.
• #31: We can summarize the findings from our experiment as follows: 1- Shared-space techniques (SG and BG) were faster than split space techniques for the local Maximum task (confirming H1) 2- On the otherhand Split-space techniques (SM and HG) were faster than shared space techniques for the dispersed Discrimination task (confirming H2) 3- The Slope task, with dispersed visual span, had a mix trend —SM and SG were fastest here 4- Higher numbers of concurrent time series caused decreased correctness and increased completion time (confirming H3) 5- Decreased display space allocation had a negative impact on correctness, but had a little effect on time (partially confirming H4).
• #32: [click] I have presented results from a user study on the graphical perception of multiple time series [click] We found that superimposed/shared space line graph techniques work best for local tasks, whereas juxtaposed/split space techniques work best for dispersed ones