<-- Advertise here.

Data visualization consultant Stephen Few has just posted an excellent article [perceptualedge.com] about a recent time-series visualization called the Horizon Graph. Originally developed by data visualization software firm Panopticon, Horizon Graphs can display about 50 sets of time-series values on a single screen. This particular visualization technique was the focus of a detailed evaluation study described by Jeffrey Heer, graduate researcher Nicholas Kong, and Maneesh Agrawala. The results are described in the scientific paper titled "Sizing the Horizon: The Effects of Chart Size and Layering on the Graphical Perception of Time Series Visualizations" (PDF) [berkeley.edu].

In the paper, 2 different experiments are described. The goal of the first experiment was to determine the impact of the band number and horizon graph variant ("mirrored", flipping the negative values around zero, versus "offset") on value comparisons between horizon graphs. The goal of the second experiment was to compare normal line charts to horizon graphs and investigate the effect of chart height on both.

Want to know the dry results? No significant difference was found in either estimation time or accuracy between the different chart types. However, both estimation time and error increased as more bands were displayed. In the 2nd experiment, the estimation error increased as chart size decreased; layering increased estimation time, and mirroring did not; and lastly, the estimation time decreases with chart height.

More interestingly, from the results, following 3 design guidelines were proposed:

  • Mirroring Does Not Hamper Graphical Perception. Mirroring a chart (that is flipping the negative values around zero) neither slowed estimation time nor hurt estimation accuracy, but cuts the size of the chart in half.

  • Layered Bands Are Beneficial As Chart Size Decreases. Dividing a chart into layered bands increased the estimation time and increased the estimation error at constant chart heights. Therefore, the use of 4 or more bands is discouraged, as this resulted in increased time and error, while subjects complained that interpreting 4-band charts was difficult and tiring

  • Optimal Chart Sizing. For both normal line charts and 1-band mirror charts, they found a chart height of 24 pixels to be optimal.

Via Datavisualization.

UPDATE: Jeffrey Heer pointed out some details that might be misinterpreted: "We do recommend a height of 24 pixels, but this is assuming a 14.1 inch 1024 x 768 display. Given the wide variety of display types and resolution, the size in millimeters (6.8mm) is likely a more informative measure for others to adopt." Thanks for the great research, Jeffrey!




At first I was confused on how to read this chart, then I thought it would be hard to understand that not only the height indicates "good or bad" but the color shows the direction of the amplitude. But as I tried to generally get an overview I realized the reading was rather easy after I learned how to read it. And the adoption time is astonishingly short.

I'd love to see a comparison between two variations of the visualization: one using the normal and the other the mirrored position of negative values. I expect negatives coming from the top could be destracting from the essential information.

btw: Thanks for the credits, but we'd like to direct'em straight to Stephen Few.

Tue 20 Jan 2009 at 7:44 AM

It is great for us at Panopticon to see this kind of pick-up on the article from Heer et al, and Stephen Few’s blog post about it. I would like to let all readers know that software developers at corporations or academic institutions are welcome to download a 30-day evaluation copy of our software. This will allow you to try out the Horizon Graph first hand. We have identical offerings for both Java and .NET.

Use the request form on this page, and specify either “Panopticon Developer .NET SDK” or “Panopticon Developer Java SDK”:


Tue 20 Jan 2009 at 8:24 PM

The measurement of the effects of changing visualisation is very interesting. I wonder if it's possible to measure softer things such as user-comfort or user-demand/request for new visualisations?

I also starting thinking (because it's called a horizon graph), that predictive data that becomes less certain as we move forward in time could become more transparent or blurry/fuzzy. What are the benefits/drawbacks of doing this and can they be measured?

The same thing could be applied to historical data if we look back into uncertain data.


Thu 22 Jan 2009 at 9:16 AM
Matt Mapleston

This seems to me to be an extension of Edward Tufte's sparklines...very information dense and still easy to comprehend!

Wed 25 Mar 2009 at 2:29 AM
Commenting has been temporarily disabled.