Graphics Reference
In-Depth Information
in which various views of the same dataset are combined in one window. One plot -
typically a scatterplot matrix - controls which variables are displayed in all the other
plots includedinthewindow.Clicking acellin thescatterplot matrix showsmarginal
views for these variables, for example. he aim of such a system - called empirical
linking in Young et al. (
) - is to create multiple views that act as a single visu-
alization of the data space and provide a complete picture by offering a magnitude
of view points and visual aspects. he major benefit lies in the fact that the user can
routinize and partially automatize the exploration of a dataset. Once interesting and
appropriate plots have been found for one variable, the user can investigate similar
or related variables from the same angles by simply exchanging the variables in the
control plot.
he model layer of a data display is flexible enough to comprise also more com-
plex models such as regression models, grand tours, and principal components. For
these models, a straightforward model link consists of a system of intertwined plots
that display the raw observations, the model, and residual information. Young et al.
(
) had introduced such a form in the context of grand tour plots. hey designed
a spread plot consisting of a rotating plot and two scatterplots presenting residual
information for the main components. Similarly, in Manet each biplot is accompa-
nied by two scatterplots of residuals showing the residuals against the first and sec-
ondprincipalcomponent;seeHofmann(
).Changesinthemodel,asforexample
initiated by rotating the point cloud in a spread plot, results in an immediate update
of the residual information. Young et al. (
) called this form of linking algebraic
linking.
Linking Types
8.2.3
he type layer covers most of the visible components in a graphical display and aims
at representing the model as well as possible. he distinction between type level and
model level basically lies in the fact that, due to the limited plot space and screen
resolution, not all models can be visualized without loss of information. he close
connection between the two layers also means that congruities at the type level of
twodisplaysalmostalwaysisaconsequenceoflinkedmodels.Forexample,itisclear
that two histograms that usethe same categorization operator also have the same bin
widths. A direct link between the type levels of two displays without having a cor-
responding linkage between the models is rather uncommon. Color and size are at-
tributes of graphical elements that can belinked, inmostcases either with orwithout
establishing a corresponding modellink. Pie charts, for example, use different colors
for the slices to enhance the differentiation between the various categories. hese
colors are typically assigned individually for each plot and do not carry any intrin-
sic meaning. Colors could be linked on a type level by assigning identical colors to
the first slice in each plot, the second slice, and so on. As long as the ordering of the
slices does not reflect information of the model level, the linking will be on the type
level only. Alternatively, if the slices are ordered alphabetically or by size, the color
also carries some model information and linking via colors could be a realization of
