i
i
i
i
i
i
i
i
27.6. Composite and Adjacent Views 695
most clearly. The power of linked highlighting across multiple visual encodings
is that items that fall in a contiguous region in one view are often distributed very
differently in the other views. In the small-multiples approach, each view has
the same visual encoding for different datasets, usually with shared axes between
frames so that comparison of spatial position between them is meaningful. Side-
by-side comparison with small multiples is an alternative to the visual clutter of
superimposing all the data in the same view, and to the human memory limitations
of remembering previously seen frames in an animation that changes over time.
The overview-and-detail approachis to have the same data and the same visual
encoding in two views, where the only difference between them is the level of
zooming. In most cases, the overview uses much less display space than the
detail view. The combination of overview and detail views is common outside
of visualization in many tools ranging from mapping software to photo editing.
With a detail-on-demand approach, another view shows more information about
some selected item, either as a popup window near the cursor or in a permanent
window in another part of the display.
Figure 27.16. The Improvise toolkit was used to create this multiple-view visualization.
Image courtesy Chris Weaver.
(See also Plate XLVIII.)
i
i
i
i
i
i
i
i
696 27. Visualization
Determining the most appropriate spatial position of the views themselves
with respect to each other can be as significant a problem as determining the
spatial position of marks within a single view. In some systems, the location of the
views is arbitrary and left up to the window system or the user. Aligning the views
allows precise comparison between them, either vertically, horizontally, or with
an array for both directions. Just as items can be sorted within a view, views can
be sorted within a display, typically with respect to a derived variable measuring
some aspect of the entire view as opposed to an individual item within it.
Figure 27.16 shows a visualization of census data that uses many views. In
addition to geographic information, the demographic information for each county
includes population, density, gender, median age, percent change since 1990,
and proportions of major ethnic groups. The visual encodings used include ge-
ographic, scatterplot, parallel coordinate, tabular, and matrix views. The same
color encoding is used across all the views, with a legend in the bottom mid-
dle. The scatterplot matrix shows linked highlighting across all views, where
the blue items are close together in some views and scattered in others. The
map in the upper-left corner is an overview for the large detail map in the cen-
ter. The tabular views allow direct sorting by and selection within a dimension
of interest.
27.7 Data Reduction
The visual encoding techniques that we have discussed so far show all of the items
in a dataset. However, many datasets are so large that showing everything simul-
taneously would result in so much visual clutter that the visual representation
would be difficult or impossible for a viewer to understand. The main strategies
to reduce the amount of data shown are overviews and aggregation, filtering and
navigation, the focus+context techniques, and dimensionality reduction.
27.7.1 Overviews and Aggregation
With tiny datasets, a visual encoding can easily show all data dimensions for all
items. For datasets of medium size, an overview that shows information about
all items can be constructed by showing less detail for each item. Many datasets
have internal or derivable structure at multiple scales. In these cases, a multiscale
visual representation can provide manylevels of overview, rather than just a single
level. Overviews are typically used as a starting point to give users clues about
where to drill down to inspect in more detail.
i
i
i
i
i
i
i
i
27.7. Data Reduction 697
For larger datasets, creating an overview requires some kind of visual sum-
marization. One approach to data reduction is to use an aggregate representation
where a single visual mark in the overview explicitly represents many items.
The challenge of aggregation is to avoid eliminating the interesting signals
in the dataset in the process of summarization. In the cartographic literature, the
problem of creating maps at different scales while retaining the important dis-
tinguishing characteristics has been extensively studied under the name of carto-
graphic generalization (Slocum et al., 2008).
27.7.2 Filtering and Navigation
Another approach to data reduction is to filter the data, showing only a subset of
the items. Filtering is often carried out by directly selecting ranges of interest in
one or more of the data dimensions.
Navigation is a specifickindoffiltering based on spatial position, where
changing the viewpoint changes the visible set of items. Both geometric and non-
geometric zooming are used in visualization. With geometric zooming, the cam-
era position in 2D or 3D space can be changed with standard computer graphics
controls. In a realistic scene, items should be drawn at a size that depends on their
distance from the camera, and only their apparent size changes based on that dis-
tance. However, in a visual encoding of an abstract space, nongeometric zooming
can be useful. In semantic zooming, the visual appearance of an object changes
dramatically based on the number of pixels available to draw it. For instance, an
abstract visual representation of a text file could change from a tiny color-coded
box with no label to a medium-sized box containing only the filename as a text
label to a large rectangle containing a multi-line summary of the file contents. In
realistic scenes, objects that are sufficiently far away from the camera are not vis-
ible in the images, for example, after they subtend less than one pixel of screen
area. With guaranteed visibility, one of the original or derived data dimensions is
used as a measure of importance, and objects of sufficient importance must have
some kind of representation visible in the image plane at all times.
27.7.3 Focus+Context
Focus+context techniques are another approach to data reduction. A subset of the
dataset items are interactively chosen by the user to be the focus and are drawn
in detail. The visual encoding also includes information about some or all of the
rest of the dataset shown for context, integrated into the same view that shows the
i
i
i
i
i
i
i
i
698 27. Visualization
focus items. Many of these techniques use carefully chosen distortion to combine
magnified focus regions and minified context regions into a unified view.
One common interaction metaphor is a moveable fisheye lens. Hyperbolic
geometry provides an elegant mathematical framework for a single radial lens
that affects all objects in the view. Another interaction metaphor is to use mul-
tiple lenses of different shapes and magnification levels that affect only local re-
gions. Stretch and squish navigation uses the interaction metaphor of a rubber
sheet where stretching one region squishes the rest, as shown in Figure 27.17.
The borders of the sheet stay fixed so that all items are within the viewport, al-
though many items may be compressed to subpixel size. The fisheye metaphor
is not limited to a geometric lens used after spatial layout; it can be used directly
on structured data, such as a hierarchical document where some sections are col-
lapsed while others are left expanded.
These distortion-based approaches are another example of non-literal navi-
gation in the same spirit as nongeometric zooming. When navigating within a
large and unfamiliar dataset with realistic camera motion, users can become dis-
oriented at high zoom levels when they can see only a small local region. These
approaches are designed to provide more contextual information than a single
Figure 27.17. The TreeJuxtaposer system features stretch and squish navigation and guar-
anteed visibility of regions marked with colors (Munzner et al., 2003). (See also Plate XLIX).
i
i
i
i
i
i
i
i
27.7. Data Reduction 699
Figure 27.18. The SpaceTree system shows the path between the root and the interactively
chosen focus node to provide context (Grosjean et al., 2002).
undistorted view, in hopes that people can stay oriented if landmarks remain rec-
ognizeable. However, these kinds of distortion can still be confusing or difficult
to follow for users. The costs and benefits of distortion, as opposed to multiple
views or a single realistic view, are not yet fully understood. Standard 3D per-
spective is a particularly familiar kind of distortion and was explicitly used as a
form of focus+context in early visualization work. However, as the costs of 3D
spatial layout discussed in Section 27.4 became more understood, this approach
became less popular.
Other approaches to providing context around focus items do not require dis-
tortion. For instance, the SpaceTree system shown in Figure 27.18 elides most
nodes in the tree, showing the path between the interactively chosen focus node
and the root of the tree for context.
27.7.4 Dimensionality Reduction
The data reduction approaches covered so far reduce the number of items to
draw. When there are many data dimensions, dimensionality reduction can also be
effective.
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.