Geoscience Reference
In-Depth Information
insight to gain.* Of course, various other filters - both statistical and geographical - may be used
so that only some attributes are graphed or certain ranges of values are plotted (Weaver, 2010).
Small multiples are useful when several possible explanations are being pursued, for example, a
systematic search of disease covariates to find those that seem to have the strongest geographical
clustering. In Figure 5.5, a matrix of six variables displayed as scatterplots, bivariate maps and
histograms (on the diagonal of the matrix) is shown in the upper left. This compound display
simultaneously supports pairwise exploration for correlation and spatial association across a sig-
nificant amount of data.
Small multiples scale up well to around 10 or so attributes on today's computer screens, after
which further reduction is necessary to see comparisons across all data attributes concurrently. To
achieve this reduction, we need to lose some detail. One approach is to simplify the matrix to only
simple values of correlation or conditional entropy, using hue or brightness to show the strength of
the result. This approach produces large numbers of
tiles
, all of which look basically similar, so
it can be daunting to use. But it does highlight sub-spaces in the data where further investigation
might be warranted (Figure 5.6).
Some geographical processes and objects can be difficult to identify because they are a result of
subtle relationships over time, space and many dimensions of attributes. That being the case, they
may not be easily discovered when these correlations are distributed between different bivariate dis-
plays. To be noticed, they may require the evidence to be combined from several data attributes con-
currently. So rather than visualise pairwise data values and their patterns, we may instead construct
a scene that combines several different data attributes directly, by creating more complex glyphs
and layers that provide more empty slots for visual variables that we can assign data to. This is eas-
ily achieved by utilising the greater visual control that fully rendered systems provide (see earlier
FIGURE 5.6
A display showing the pairwise correlation (upper right diagonal) and entropy (lower right
diagonal) across a dataset containing 36-variable demographic and health-related data attributes. Each tile
in the display represents the entropy or correlation between a pair of data attributes. Although not visible
from the above figure, red values shown within the display indicate more similarity between the attributes.
Clusters of colours show regions in this combined dataset that might warrant more detailed investigation,
perhaps using some of the display types shown earlier in Figure 5.5.
*
From the discussion of the science discovery process presented later we could perhaps claim that the hypothesis space for
discovery is evenly sampled, without initial bias in this starting configuration.
