Graphics Reference
In-Depth Information
Figure
.
.
hree example screenshots of three different projections of the cars data set. he cases are
colored according to the number of cylinders. he rightmost plot has the least discrimination of the
groups but the strongest separation of an outlier, the “Porsche GT”
theprojections arenotverysatisfactory unlesstheyreveal astriking feature.Whereas
the parallel coordinate plot of the same data (cf. Fig.
.
) can at least show the uni-
variate distributions along with some bivariate relationships, the grand tour fails to
doso, and focuses solely on the multivariate features, whichmaybe visible in certain
projections.
he grand tour can help to identify the geometry of variables beyond the limits of
the three dimensions of a simple rotating plot. Nevertheless, examples of structures
inmorethan five dimensions are rare, even whenusing the grand tour. Inthese cases
the fixed geometrical propertiesof parallel coordinate plots seem tobean advantage.
Monitoring the projected data in parallel coordinates instead of a simple scatter-
plot is a promising approach to investigating data beyond ten or even more dimen-
sions. Unfortunately, only very few flexible implementations of the grand tour and
projection pursuit exist, which limits the possibility of a successful application of
these methods.
Recommendations
6.6
his chapter showed the application, strengths, and weaknesses of the most impor-
tant high-dimensional plots in statistical data visualization. All plots have their spe-
cificfieldofapplication, wherenoothermethoddeliversequivalent results.hefields
of application are broader or narrower depending on the method. All four methods
andtechniques discussedinthischapter,i.e.,mosaicplots, trellisdisplays,parallelco-
ordinate plots, and the grand tour and projection pursuit, need a certain amount of
training to be effective. Furthermore, training is required to learn the most effective
use of the different methods for different tasks.
