How to visualize data with extreme value differences?

Question

I am working on some routines for a client application to visualize data in a 3d bar chart style. The data consists mostly of smaller values with only a few large values. For example:

6,942,535,341
23,598
19,203
58,201

So, the problem is that the large values pretty much makes the visualization useless. Does anyone have any suggestions on how to display this data … OR … perhaps a suggestion on how to massage the data to make it more visually appealing?

Answer 1

For anyone mathematically inclined, the answer is to use a log scale.

For non-mathematical people, you may be better off showing a break in the chart and then the extreme value.

Answer 2

You could use logarithmic axes. This allows you to compactly visualize wide ranging variables.

To illustrate, here is a very simple logarithmic visualization:

6 942 535 341
       23 598 
      419 203 
        8 201
            3

The length of each datum represented as a number is (roughly) log_10 of that number. So just printing the numbers in a monospaced font will give you a kind of automatic bar chart.

Answer 3

John GB's solution is aesthetically pleasing, but scientific fields generally discourage scale breaks like those shown (and are often cliche examples of misleading graphics). Cleveland (1984) suggests to use a full scale break, making a separate panel, to further visually distinguish between the values.

Jon Peltier has a similar example using multi-panels, but has a gradient effect for the bars going into the second higher value panel.

Scale breaks IMO are probably not as bad as everyone seems to make them, but here is a simple alternative if someone is giving you a problem with it (in addition to the other fine answers already)!

Answer 4

Get one more dimension ! Represent the values as areas. Squares or discs. Like on the anamorphic maps.

Then, if you still need more to represent your values, make cubes, or balls ! :-) 

Edit:

They use this technique in the Billion Dollar Gram.

Answer 5

Logarithmic scales are probably the best bet, though they still end up throwing off the true vastness of the difference. It really depends on what needs to be conveyed by the way that the information is displayed. If the vastness of the difference is of key importance then simply graphing the small number as a very small line may still be appropriate.

Answer 6

I think of two possible approaches:

1- Use areas. Like, a small square dot against a big rectangle area. The advantage of this approach is that the areas are proportional to the square of the measure.

2- Allow zooming. Draw the columns in scale and let the user zoom in until the small values are visible, or zoom out until the big value fits into the window.

Answer 7

I like the idea of using spherical objects where applicable. The larger areas can clearly hold several times more area while keeping the smaller areas visible.

^{download bmml source – Wireframes created with Balsamiq Mockups}

Answer 8

Log scales are the natural answer to this problem, but as mentioned above they can be misleading for the common user. A good solution to this can be to incrementally show the data points and use animations to guide the user through the understanding of the chart.

To illustrate what I mean you can take look at this visualisation from the NY Times.

These are the techniques I would highlight:

1. You load the page and an animation drops part of the dataset in place. Notice how the higher values (in this case the Facebook IPO data) are left off for the moment. Scale is linear at the moment, offering a more natural reading.

2. The user clicks "next", Facebook data comes into place. We perceive the striking difference. Scale is still linear.

3. The user clicks "next" once again and the scale becomes logarithmic (there a brief text explaining whats happening too) offering another way to read the data. Now the user can drill more easily into the finer details without losing focus on the actual size of the extremes of the data.

Answer 9

Well, I was struggling with this problem. I don't recommend to use log scale, first not all users will understand this scale, second one, if you have values that are close the difference between them won't be noticeable on the graph.

I recommend two solutions:

1. break graphs: but we need to remember to change the scale, scale must be adjusted not to the biggest value, but to the rest smaller values, for the user it's more important to see even subtle differences between smaller values, also bigger bar are easier to be clicked

2. spheres - also good solution, spehers shouldn't be proportional, cause if they are, smaller values won't be displayed - they will be so small that only one biggest value would be noticeable

Answer 10

I'd suggest providing a control to allow moving the viewpoint close enough to emphasise the smallest differences, and far enough back to bring the most extreme value fully into view. The expansion and contraction need only be on the Y axis since the X axis doesn't change.

Discontinuous scales are like log scales: it takes a lot of cognitive effort to avoid perceiving them as linear. It's better not to require that level of effort.

Answer 11

I recently encountered a situation where I not only had data with huge disparities in values. Complicating the matter further was that the data was hierarchical in nature and there was a story that needed to be shown within the second and third levels.

I found several visualizations that could at least show a flat line or teeny tiny square for the really small values for any 1 dimension.

I finally tried, and loved, the way a Sankey resolved my issue. I can see the ratio of values at any level in the hierarchy, while also seeing the big picture.