ABSTRACT
Treemaps support visualization of large hierarchical
information spaces. The treemap generation algorithm is straightforward
and application prototypes have only minimal hardware requirements.
Given primary graphical encodings of area, color and enclosure,
treemaps are best suited for the tasks of outlier detection, cause-effect
analysis and location of specific nodesósatisfying user-specified
criteriaóin their hierarchical context. Distortion effects
extend treemap capabilities by emphasizing node relationships
in the diagram.
KEYWORDS: Visualization,
statistics, hierarchy, treemap
INTRODUCTION
The large information spaces of today can only be
harnessed by new and innovative visualization tools that provide
overview, exploration and dissection capabilities. Current software
for data exploration of even moderately-sized spaces falls short
in providing this data-harnessing functionality; users are asked
to keep information in mind, use navigational tools such as scrollbars
to view small chunks of data at a time, and piece together a data
space bit by bit as a consequence.
Treemaps were developed to manage large hierarchical
information spaces without requiring workstation-class hardware;
prototypes have been implemented on a variety of platforms including
386 class DOS machines, Macintosh 68030 machines and Sun workstations.
Treemaps are generated via a simple recursive slicing algorithm
that partitions a rectangular screen area using a numeric weighting
attribute; the direction of the slice is reversed at each tree
level [2,3]. All nodes map to individual rectangles on the treemap;
rectangular enclosure is used to convey parent-child relationships
(Figure 1). As the algorithm uses all of the provided screen space,
the treemap can display about an order of
magnitude more nodes compared to traditional methods such as tree
diagrams; exact amounts depend on both screen resolutions and
the statistical distribution of the weighting attribute used by
the treemap algorithm.
Figure 1: The four NBA divisions are displayed, sliced
into teams and then players. The weighting attribute is points
per season.
Figure 2: Chicagoís Michael Jordan is emerging with the greatest area in this distorted view of Figure 1.
Many enhancements, outlined in the accompanying video
and [5], have been added to the treemap since its inception. As
the true benefits of the diagram can only be appreciated when
the domain is familiar or of personal importance, the video explores
the familiar domain of NBA basketball player statistics (1991-92
season data were used with 48 numeric attributes for over 450
players in the league).
Spreadsheets have traditionally been used to analyze
this type of data with tables upon tables of different and intriguing
statistics. Digesting a table of over 450 players is a difficult
task, however; treemaps provide a solution to this problem through
their visualization capabilities.
GRAPHICAL PROPERTIES AND TASKS
Area, color and enclosure are the inherent graphical
properties of treemaps that directly convey information. Although
an overall consensus has not been achieved as to the efficacy
and general task applicability of these properties, some guidelines
have emerged as to their usefulness when applied to particular
task domains.
Area
Area is both an asset and a liabilityótreemaps
use area to weight individual nodes, yet by the algorithmís
nature, the aspect-ratios of the generated rectangles are different,
making them ill-suited for simple comparison tasks [1,4]. Area,
though, does have its benefits: because an overview of the entire
information space is provided, the treemap allows the user to
perform outlier identification tasks based upon the areas of individual
rectangles. In the NBA example, a weighting attribute of ìpoints
per seasonî translates into large areas for players who
have achieved high point totals over the entire season (this attribute
can be inverted to identify low-scoring players as well).
Relative comparisons of siblings (teams within a
division, players of a team) can also be accomplished as all children
of a node are either the same height or width; viewed in the context
of this task, the treemap becomes a meta-chart of relative bar
charts.
Color
Color is also used to convey attributes, which may
be continuous or categorical. For the NBA domain, color was applied
to all of the numeric attributes used for area weightingóspecific
class intervals were established based on these attributesí
maxima, minima and data distribution.
Careful consideration should be given to the use
of color with treemaps given the potential for interaction effects.
A more efficient use of color is as a filtering mechanism: the
video demonstrates color as a highlighting tool for nodes that
satisfy certain criteria. This accomplishes the task of pinpointing
the location of specific nodes in the context of the complete
hierarchy.
Enclosure
The concept of enclosure to indicate parentage has
been well-studied and applied in many different domains. The enclosure
provided by the treemaps triggers location-oriented tasks: are
all high-scoring players located under a particular division or
team? Enclosure also emphasizes cause-effect relationships. Child
node weighting influences parent node weighting; sibling nodes
compete for each otherís space. Large child nodes, therefore,
create a ìrippleî effect up the hierarchy making
cause-effect relationships obvious.
AREA DISTORTION
Difficulty arises when there is low variance in the
underlying data points used as the weighting attribute. For this
case, as well as enhancing the comparison capabilities of the
diagram, distortion techniques have been introduced. Visually
altering the rectangular areas through distortion clarifies node
relationships in the diagram. Treemap distortion can be accomplished
in three ways: altering the underlying weights that the treemap
algorithm uses, applying geometric transformations to the diagram,
or allowing the user to directly manipulate the treemap.
The first technique was used in the video. Each playerís
weight was altered using an exponential function. The visual impact
of this function is that large areas grow even larger, overwhelming
their smaller siblings and cousins; the effect is quite similar
to a fisheye diagram with multiple foci. Figure 2 illustrates
distortion on Figure 1; players with large season point totals
have larger areas. Each of the four divisions is seen to have
one or more standout players.
More advanced algorithms for providing distortion
fluidity and efficiency are under research.
REFERENCES
1. Cleveland, W.S. and McGill, R. Graphical Perception:
Theory, Experimentation and Application to the Development of
Graphical Methods. Journal of the American Statistical Association
79, 387, 1984, pp. 531-554.
2. Johnson, B. and Shneiderman, B. Tree-Maps: A Space-Filling
Approach to the Visualization of Hierarchical Information Structures.
In Proceedings of IEEE Visualizationí91 (Oct. 22-25,
San Diego, CA), IEEE Computer Society Press, 1991, pp. 284-291.
3. Shneiderman, B. Tree Visualization with Tree-maps:
A 2-D space-filling approach. ACM Transactions on Graphics
11, 1 (Jan. 1992), pp. 92-99.
4. Simkin, D. and Hastie, R. An Information-Processing
Analysis of Graph Perception. Journal of the American Statistical
Association 82, 398, 1987, pp. 454-465.
5. Turo, D. and Johnson, B. Improving the Visualization
of Hierarchies with Treemaps: Design Issues and Experimentation.
In Proceedings of IEEE Visualizationí92 (Oct. 8-12,
Boston, MA), IEEE Computer Society Press, 1992, pp. 124-131.