See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/311706649

Information Visualization

Chapter · January 2008

CITATIONS READS
0 13,918

1 author:

Stuart K. Card
Stanford University
206 PUBLICATIONS 32,440 CITATIONS

SEE PROFILE

All content following this page was uploaded by Stuart K. Card on 17 December 2016.

The user has requested enhancement of the downloaded file.

 ◆
26 ◆

INFORMATION VISUALIZATION

Stuart Card
Xerox PARC

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510 Double-axis composition . . . . . . . . . . . . . . . . . . . .532

Example 1: Finding Videos with the FilmFinder . . . . . . 510 Mark composition and case composition . . . . . . . .533
Example 2: Monitoring Stocks with TreeMaps . . . . . . . 512 Recursive composition . . . . . . . . . . . . . . . . . . . . . .533
Example 3: Sensemaking with Permutation Interactive Visual Structures . . . . . . . . . . . . . . . . . . . . . . 534
Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513 Dynamic queries . . . . . . . . . . . . . . . . . . . . . . . . . . .534
What Is Information Visualization? . . . . . . . . . . . . . . . . . 515 Magic lens (movable filter) . . . . . . . . . . . . . . . . . . .534
Why Does Visualization Work? . . . . . . . . . . . . . . . . . . . . . 515 Overview 1 detail . . . . . . . . . . . . . . . . . . . . . . . . . .534
Historical Origins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516 Linking and brushing . . . . . . . . . . . . . . . . . . . . . . .534
The Visualization Reference Model . . . . . . . . . . . . . . . . 519 Extraction and comparison . . . . . . . . . . . . . . . . . . .536
Mapping Data to Visual Form . . . . . . . . . . . . . . . . . . . . . 519 Attribute explorer . . . . . . . . . . . . . . . . . . . . . . . . . .536
Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519 Focus 1 Context Attention-Reactive
Visual Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520 Abstractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
Spatial substrate . . . . . . . . . . . . . . . . . . . . . . . . . . .520 Data-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
Marks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .521 Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .536
Connection and enclosure . . . . . . . . . . . . . . . . . . .522 Selective aggregation . . . . . . . . . . . . . . . . . . . . . . .537
Retinal properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522 View-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
Temporal encoding . . . . . . . . . . . . . . . . . . . . . . . . .522 Micro-macro readings . . . . . . . . . . . . . . . . . . . . . . . . 537
Expressiveness and Effectiveness . . . . . . . . . . . . . . . . . . 523 Highlighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .537
Taxonomy of Information Visualizations . . . . . . . . . . . . 523 Visual transfer functions . . . . . . . . . . . . . . . . . . . . .537
Simple Visual Structures . . . . . . . . . . . . . . . . . . . . . . . . . 524 Perspective distortion . . . . . . . . . . . . . . . . . . . . . . .538
1-Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525 Alternate geometries . . . . . . . . . . . . . . . . . . . . . . . .539
2-Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526 Sensemaking With Visualization . . . . . . . . . . . . . . . . . . 539
3-Variables and Information Landscapes . . . . . . . . . . . . 526 Knowledge Crystallization . . . . . . . . . . . . . . . . . . . . . . . . 539
n-Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527 Acquire information . . . . . . . . . . . . . . . . . . . . . . . .540
Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527 Make sense of it . . . . . . . . . . . . . . . . . . . . . . . . . . . .540
Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .527 Create something new . . . . . . . . . . . . . . . . . . . . . .540
Enclosure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530 Act on it . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .540
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530 Levels for Applying Information Visualization . . . . . . . . 540
Composed Visual Structures . . . . . . . . . . . . . . . . . . . . . . 531 Acknowledgment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 542
Single-axis composition . . . . . . . . . . . . . . . . . . . . .531 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 542

509
510 • CARD

1994b). Unlike typical movie-finder systems, the FilmFinder is

INTRODUCTION organized not around searching with keywords, but rather
around rapid browsing and reacting to collections of films in the
The working mind is greatly leveraged by interaction with the database Figure 26.1 shows a scattergraph of 2000 movies, plot- [AQ1]
world outside it. A conversation to share information, a grocery ting rated quality of the movie as a function of year when it was
list to aid memory, a pocket calculator to compute square released. Color differentiates type of movies—comedy from
roots—all effectively augment a cognitive ability otherwise se- drama and the like. The display provides an overview, the en-
verely constrained by what is in its limited knowledge, by lim- tire universe of all the movies, and some general features of the
ited attention, and by limitations on reasoning. But the most collection. It is visually apparent, for example, that a good share
profound leverage on cognitive ability is the ability to invent of the movies in the collection were released after 1965, but also
new representations, procedures, or devices that augment cog- that there are movies going back as far as the 1920s. Now the
nition far beyond its unaided biological endowment—and boot- viewer “drills down” into the collection by using the sliders in
strap these into even more potent inventions. the interface to show only movies with Sean Connery that are
This chapter is about one class of inventions for augmenting between 1 and 41⁄2 hours in length (Fig. 26.2). As the sliders are [AQ2]
cognition, collectively called “information visualization.” Other moved, the display zooms in to show about 20 movies. It can
senses could be employed in this pursuit—audition, for exam- be seen that these movies were made between 1960 and 1995,
ple, or a multi-modal combination of senses—the broader topic and all have a quality rating higher than 4. Since there is now
is really information perceptualization; however, in this chap- room on the display, titles of the movies appear. Experimenta-
ter, we restrict ourselves to visualization. Visualization employs tion with the slider shows that restricting maximum length to
the sense with the most information capacity; recent advances in 2 hours cuts out few interesting movies. The viewer chooses the
graphically agile computers have opened opportunities to ex- highly rated movie, “Murder on the Orient Express” by double-
ploit this capacity, and many visualization techniques have now clicking on its marker. Up pop details in a box (Fig. 26.3) giving [AQ3]
been developed. A few examples suggest the possibilities. names of other actors in the movie and more information. The
viewer is interested in whether two of these actors, Anthony
Perkins and Ingrid Bergman, have appeared together in any
Example 1: Finding Videos with the FilmFinder other movies. The viewer selects their names in the box, and
then requests another search (Fig. 26.4). The result is a new dis- [AQ4]
The use of information visualization for finding things is illus- play of two movies. In addition to the movie the viewer knew
trated by the FilmFinder (Ahlberg & Shneiderman, 1994a, about, there is one other movie, a drama entitled “Goodbye,

FIGURE 26.1. FilmFinder overview scattergraph. Courtesy University of Maryland.

FIGURE 26.2. FilmFinder scattergraph zoom-in. Courtesy University of Maryland.

FIGURE 26.3. FilmFinder details on demand. Courtesy University of Maryland.

511
 512 • CARD

FIGURE 26.4. FilmFinder retrieval by example. Courtesy University of Maryland.

Again,” made around 1960. The viewer is curious about this more than 500 stocks at once, with data updated every 15 min-
movie and decides to watch it. utes. Each colored rectangle in the figure is a company. The size
Information visualization has allowed a movie viewer in a of the rectangle is proportional to its market capitalization.
matter of seconds to find a movie he or she could not have spec- Color of the rectangle shows movement in the stock price.
ified at the outset. To do this, the FilmFinder employed several Bright yellow corresponds to about a 6% increase in price,
techniques from information visualization: (a) an overview of bright blue to about a 6% decrease in price. Each business sec-
the collection showing its structure; (b) dynamic queries, in tor is identified with a label like “Communications.” Those items
which the visualization seems to change instantaneously with marked with a letter N have an associated news item.
control manipulations; (c) zooming in by adding restrictions In this example, the investor’s task is to monitor the day’s
to the set of interest; (d) details on demand, in which the user market and notice interesting developments. In Fig. 26.5(a), the
can display temporarily detail about an individual object, and investor has moved the mouse over one of the bright yellow
(e) retrieval by example, in which selected attributes of an in- rectangles, and a box identifying it as Erickson, with a 9.28%
dividual object are used to specify a new retrieval set. gain for the day, has popped up together with other informa-
tion. Clicking on a box gives the investor a popup menu for se-
lecting even more detail. The investor can either click to go to
Example 2: Monitoring Stocks with TreeMaps World Wide Web links on news or financials, or drill down, for [AQ7]
example, to the sector (Fig. 26.5[b]), or down further to indi-
Another example of information visualization is the TreeMap vi- vidual companies in the software part of the technology sector
[AQ5] sualization on the SmartMoney.com website,1 which is shown in (Fig. 26.5[c]). The investor is now able to immediately note in-
[AQ6] Fig. 26.5(a). Using this visualization, an investor can monitor teresting relationships. The software industry is now larger than

[AQ55] 1
www.smartmoney.com
 26. Information Visualization • 513

FIGURE 26.5. TreeMap of daily stock prices. Courtesy SmartMoney.com.

the hardware industry, for example, and despite a recent bat- Example 3: Sensemaking with Permutation Matrices
tering at the time of this figure, the Internet industry is also rel-
atively large. Microsoft is larger than all the other companies in As a final information visualization example, consider the case
its industry combined. Selecting a menu item to look at year- proposed by Bertin (1977/1981) of a hotel manager who wants
[AQ8] to-date gains (Fig. 26.6), the investor immediately notes inter- to analyze hotel occupancy data (Table 26.1) to increase her re- [AQ11]
esting patterns: Microsoft stock shows substantial gains, turn. In order to search for meaningful patterns in her data, she
whereas Oracle is down; Dell is up, but Compaq is down; Tiny represents it as a permutation matrix (Fig. 26.7[a]. A permuta-
Advanced Micro is up, whereas giant Intel is neutral. Having no- tion matrix is a graphic rendition of a cases x variables display. In
[AQ9] ticed these relationships, the investor drills down to put up Fig. 26.7(a), each cell of Table 26.1 is a small bar of a bar chart.
[AQ10] charts or analysts positions for companies whose gains in them- The bars for cells below the mean are white; those above the bar
selves, or in relation to a competitor, are interesting. For ex- are black. By permuting rows and columns, patterns emerge
ample, the investor is preparing a report on the computer in- that lead to making sense of the data.
dustry for colleagues and notices how AMD is making gains In Fig. 26.7(a), the set of months, which form the cases, are
against Intel, or how competition for the Internet is turning into repeated to reveal periodic patterns across the end of the cy-
a battle between Microsoft and AOL/Time Warner. cle. By visually comparing the pairs of rows, one can find rows
 FIGURE 26.6. TreeMap of year-to-date stock prices. Courtesy SmartMoney.com.

TABLE 26.1. Data for Hotel Occupancy (Based on Bertin (1977/1981))

ID VARIABLE JAN FEB MAR APR MAY JUNE JULY AUG SEPT OCT NOV DEC
1 % Female 26 21 26 28 20 20 20 20 20 40 15 40
2 % Local 69 70 77 71 37 36 39 39 55 60 68 72
3 % USA 7 6 3 6 23 14 19 14 9 6 8 8
4 % South America 0 0 0 0 8 6 6 4 2 12 0 0
5 % Europe 20 15 14 15 23 27 22 30 27 19 19 17
6 % M.East/Africa 1 0 0 8 6 4 6 4 2 1 0 1
7 % Asia 3 10 6 0 3 13 8 9 5 2 5 2
8 % Businessmen 78 80 85 86 85 87 70 76 87 85 87 80
9 % Tourists 22 20 15 14 15 13 30 24 13 15 13 20
10 % Direct Reservations 70 70 75 74 69 68 74 75 68 68 64 75
11 % Agency Reservations 20 18 19 17 27 27 19 19 26 27 21 15
12 % Air Crews 10 12 6 9 4 5 7 6 6 5 15 10
13 % Under 20 2 2 4 2 2 1 1 2 2 4 2 5
14 % 20–35 25 27 37 35 25 25 27 28 24 30 24 30
15 % 35–55 48 49 42 48 54 55 53 51 55 46 55 43
16 % Over 55 25 22 17 15 19 19 19 19 19 20 19 22
17 Price of rooms 163 167 166 174 152 155 145 170 157 174 165 156
18 Length of stay 1.7 1.7 1.7 1.91 1.9 2 1.54 1.6 1.73 1.82 1.66 1.44
19 % Occupancy 67 82 70 83 74 77 56 62 90 92 78 55
20 Conventions 0 0 0 1 1 1 0 0 1 1 1 1

514
 26. Information Visualization • 515

that are similar. These are reordered and grouped (Fig. 26.7[b]). ography, tends to emphasize finding a way to make visible the
By this means, it is discovered that there seem to be two pat- invisible (say, velocity of air flow) within an existing spatial
terns of yearly variation. One pattern in Fig. 26.7(b) is semian- framework. The chief problem for information visualization, in
nual, dividing the year into the cold months of October through contrast, is often finding an effective mapping between ab-
April and the warm months of May through September. The stract entities and a spatial representation. Both information
other pattern breaks the year into four distinct regions. We have visualization and scientific visualization belong to the broader
thus found the beginnings of a schema—that is, a framework field of data graphics, which is the use of abstract, nonrepre-
in terms of which we can encode the raw data and describe it sentational visual representations to amplify cognition. Data
in a more compact language. Instead of talking about the events graphics, in turn, is part of information design, which con-
of the year in terms of individual months, we can now talk in cerns itself with external representations for amplifying cog-
terms of two series of periods, the semiannual one, and the four nition. At the highest level, we could consider information de-
distinct periods. As we do so, there is a residue of information sign a part of external cognition, the uses of the external
not included as part of our descriptive language. Sensemaking world to accomplish some cognitive process. Characterizing
proceeds by the omission and recoding of information into the purpose of information visualization as amplifying cogni-
[AQ12&13] more compact form (see Resnikoff, 1989). This residue of in- tion is purposely broad. Cognition can be the process of writ-
formation may be reduced by finding a better or more articu- ing a scientific paper or shopping on the Internet for a cell
lated schema, or it may be left as noise. Beyond finding the ba- phone. Generally, it refers to the intellectual processes in
sic patterns in the data, the hotel manager wants to make sense which information is obtained, transformed, stored, retrieved,
of the data relative to a purpose: she wants to increase the oc- and used. All of these can be advanced generally by means of
cupancy of the hotel. Therefore, she has also permuted gen- external cognition, and in particular by means of information
eral indicators of activity in Fig. 26.7(b), such as % Occupancy visualization.
and Length of Stay, to the top of the diagram and put the rows
that correlate with these below them. This reveals that Conven-
tions, Businessmen, and Agency Reservations, all of which gen- Why Does Visualization Work?
erally have to do with convention business, are associated with
higher occupancy. This insight comes from the match in pat- Visualization aids cognition not because of some mystical su-
terns internal to the visualization; it also comes from noting periority of pictures over other forms of thought and commu-
why these variables might correlate as a consequence of factors nication, but rather because visualization helps the user by
external to the visualization. She also discovers that marked dif- making the world outside the mind a resource for thought in
ferences exist between the winter and summer guests during specific ways. We list six groups of these in Table 26.2 (Card [AQ15]
the slow periods. In winter, there are more local guests, women, et al., 1999): Visualization amplifies cognition by (a) increasing
and age differences. In summer, there are more foreign tourists the memory and processing resources available to the users,
and less variation in age. (b) reducing search for information, (c) using visual represen-
This visualization was useful for sensemaking on hotel oc- tations to enhance the detection of patterns, (d) enabling per-
cupancy data, but it is too complicated to communicate the ceptual inference operations, (e) using perceptual attention
high points. The hotel manager therefore creates a simplified mechanisms for monitoring, and (f ) by encoding information
diagram, Fig. 26.7(c). By graying some of the bars, the main in a manipulable medium. The FilmFinder, for example, allows
points are more readily graspable, while still preserving the the representation of a large amount of data in a small space
data relations. A December convention, for example, does not in a way that allows patterns to be perceived visually in the
seem to have the effect of the other conventions in bringing in data. Most important, the method of instantly responding in
guests. It is shown in gray as residue in the pattern. The hotel the display to the dynamic movement of the sliders allowed
manager suggests moving the convention to another month, users to rapidly explore the multidimensional space of films.
where it might have more effect on increasing the occupancy of The TreeMap of the stock market allows monitoring and explo-
the hotel. ration of many equities. Again, much data is represented in lit-
tle space. In this case, the display manages the user’s attention,
drawing it to those equities with unusually large changes, and
What Is Information Visualization? supplying the means to drill down into the data to understand
why these movements may be happening. In the hotel man-
The FilmFinder, the TreeMap, and the permutation matrix ho- agement case, the visual representation makes it easier to no-
tel analysis are all examples of the use of information visual- tice similarities of behavior in a multidimensional attribute
ization. We can define information visualization as “the use space, then to cluster and rerepresent these. The final prod-
of computer-supported, interactive, visual representations of uct is a compact (and simplified) representation of the origi-
abstract data in order to amplify cognition” (Card, Mackinlay, & nal data that supports a set of forward decisions. In all of these
Shneiderman, 1999). cases, visualization allows the user to (a) examine a large
Information visualization needs to be distinguished from amount of information, (b) keep an overview of the whole
related areas: scientific visualization is like information visu- while pursuing details, (c) keep track of (by using the display as
alization, but it is applied to scientific data and typically is an external working memory) many things, and (d) produce an
[AQ14] physically based. The starting point of a natural geometrical abstract representation of a situation through the omission and
substrate for the data, whether the human body or earth ge- recoding of information.
516 • CARD

FIGURE 26.7. Permutation matrix representation of hotel data from (Berlin, 1977/1981). (a) Initial
matrix of variables. (b) Permuted matrix to group like patterns together.

Historical Origins (a) record observations, (b) induct relationships, (c) explicate
methodology of experiments, and (d) classify and conceptualize
Drawn visual representations have a long history. Maps go back phenomena (for a discussion, see Robin, 1992). For example,
millennia. Diagrams were an important part of Euclid’s books Fig. 26.8 is a hand-drawn illustration, in Newton’s first scientific [AQ16]
on geometry. Science, from earliest times, used diagrams to publication, illustrating how white light is really composed of
 26. Information Visualization • 517

TABLE 26.2. How Information Visualization Amplifies Cognition

1. Increased Resources
High-bandwidth hierarchical interaction Human moving gaze system partitions limited channel capacity so that it combines high spatial
resolution and wide aperture in sensing the visual environments (Larkin & Simon, 1987).
Parallel perceptual processing Some attributes of visualizations can be processed in parallel compared to text, which is serial.
Offload work from cognitive to Some cognitive inferences done symbolically can be recoded into inferences done with simple
perceptual system perceptual operations (Larkin & Simon, 1987).
Expanded working memory Visualizations can expand the working memory available for solving a problem (Norman, 1993).
Expanded storage of information Visualizations can be used to store massive amounts of information in a quickly accessible form
(e.g., maps).
2. Reduced Search
Locality of processing. Visualizations group information used together reducing search (Larkin & Simon, 1987).
High data density. Visualizations can often represent a large amount of data in a small space (Tufte, 1983).
Spatially-indexed addressing. By grouping data about an object, visualizations can avoid symbolic labels (Larkin & Simon, 1987).
3. Enhanced Recognition of Patterns
Recognition instead of recall. Recognizing information generated by a visualization is easier than recalling that information by
the user.
Abstraction and aggregation Visualizations simplify and organize information, supplying higher centers with aggregated forms of
information through abstraction and selective omission (Card, Robertson, & Mackinlay, 1991);
(Resnikoff, 1989).
Visual schemata for organization Visually organizing data by structural relationships (e.g., by time) enhances patterns.
Value, relationship, trend Visualizations can be constructed to enhance patterns at all three levels (Bertin, 1967/1983).
4. Perceptual Inference
Visual representations make some
problems obvious Visualizations can support a large number of perceptual inferences that are very easy for humans
(Larkin & Simon, 1987).
Graphical computations Visualizations can enable complex specialized graphical computations (Hutchins, 1996).
5. Perceptual Monitoring Visualizations can allow for the monitoring of a large number of potential events if the display is
organized so that these stand out by appearance or motion.
6. Manipulable medium Unlike static diagrams, visualizations can allow exploration of a space of parameter values and can
amplify user operations.
Source: Card, Mackinlay, & Shneiderman, 1999.

many colors. Sunlight enters from the window at right and is presentation, but on the use of pictures to give rapid, statistical
refracted into many colors by a prism. One of these colors can insight into data relations. For example, “box and whisker plots”
be selected (by an aperture in a screen) and further refracted by allowed an analyst to get a rapid characterization of data distri-
another prism, but the light stays the same color, showing that butions. Cleveland and McGill (1988) wrote an influential book,
it has already been reduced to its elementary components. As in Dynamic Graphics for Statistics, explicating new visualizations
Newton’s illustration, early scientific and mathematical diagrams of data with particular emphasis on the visualization of multidi-
generally had a spatial, physical basis and were used to reveal mensional data.
the hidden, underlying order in that world. In 1985, NSF launched an initiative on scientific visualiza-
Surprisingly, diagrams of abstract, nonphysical information tion (McCormick & DeFanti, 1987). The purpose of this initia-
are apparently rather recent. Tufte (1983) dates abstract dia- tive was to use advances in computer graphics to create a new
[AQ17] grams to (Playfair, 1786) in the 18th century. Figure 26.9 is one class of analytical instruments for scientific analysis, especially as
of Playfair’s earliest diagrams. The purpose was to convince a tool for comprehending large, newly produced datasets in the
readers that English imports were catching up with imports. geophysical and biological sciences. Meanwhile, the computer
Starting with Playfair, the classical methods of plotting data were graphics and artificial intelligence communities were interested
developed—graphs, bar charts, and the rest. in the automatic design of visual presentations of data. Mackinlay’s
Recent advances in the visual representation of abstract in- (1986a, 1986b) thesis APT formalized Bertin’s design theory,
formation derive from several strands that became intertwined. added psychophysical data, and used these to build a system for
In 1967, Bertin (1967/1983, 1977/1981), a French cartographer automatically generating diagrams of data, tailored for some
published his theory of The Semiology of Graphics. This theory purpose. Roth and Mattis (1990) built a system to do more com-
[AQ18] identified the basic elements of diagrams and their combina- plex visualizations, such as some of those from Tufte. Casner
tion. Tufte (1983, 1990, 1997), from the fields of visual design (1991) added a representation of tasks. This community was in-
and data graphics, published a series of seminal books that set terested not so much in the quality of the graphics as in the au-
forth principles for the design of data graphics and emphasized tomation of the match between data characteristics, presenta-
maximizing the density of useful information. Both Bertin’s and tional purpose, and graphical presentation. Finally, the user
Tufte’s theories became well known and influential. Meanwhile, interface community saw advances in graphics hardware open-
within statistics, Tukey (1977) began a movement on exploratory ing the possibility of a new generation of user interfaces. The
data analysis. His emphasis was not on the quality of graphical first use of the term “information visualization” was probably in
518 • CARD

FIGURE 26.8. Newton’s optics illustration (from Robin, 1992).

FIGURE 26.9. Playfair’s charts of English imports and exports, from (Tufte, 1983).

Robertson, Card, and Mackinlay (1989). Early studies in this tion for trees. Robertson, Card, and Mackinlay (1993) presented
community focused on user interaction with large amounts of ways of using animation and distortion to interact with large
information: Feiner and Beshers (1990) presented a method, data sets in a system called the Information Visualizer, which
worlds within worlds, for showing six-dimensional financial data used focus  context displays to nonuniformly present large
in an immersive virtual reality. Shneiderman (1992) developed a amounts of information. The emphasis for these studies was
technique called “dynamic queries” for interactively selecting on the means for cognitive amplification, rather than on the
subsets of data items and TreeMaps, a space-filling representa- quality of the graphics presentations.
 26. Information Visualization • 519

The remainder of this chapter will concentrate on the tech- words. Document vectors, in turn, might be reduced by multi-
niques that have been developed for mapping abstract infor- dimensional scaling to create the analytic abstraction to be vi-
mation to interactive visual form to aid some intellectual task. sualized, expressed as a Data Table of x, y, z coordinates that
The perceptual foundations of this effort are beyond the scope could be displayed. These coordinates might be transformed
of this chapter, but are covered in Ware (2000). Further details into a Visual Structure—that is, a surface on an information
on information visualization techniques are addressed in a text landscape—which is then viewed at a certain angle.
by Spence (2000). The classic papers in information visualiza- Similar final effects can be achieved by transformations at dif-
tion are collected in Card et al. (1999). ferent places in the model: When a point is deleted from the vi-
sualization, has the point been deleted from the dataset? Or is it
still in the data merely not displayed? Chi and Riedl (1998) called
this the view-value distinction, and it is an example of just one
THE VISUALIZATION REFERENCE MODEL issue where identifying the locus of a transformation using the
Visualization Reference Model helps to avoid confusion.
Mapping Data to Visual Form Information visualization is about the not just creation of vi-
sual images, but also the interaction with those images in the
Despite their seeming variability, information visualizations can service of some problem. In the Visualization Reference Model,
be systematically analyzed. Visualizations can be thought of as another set of arrows flow back from the human at the right into
adjustable mappings from data to visual form to the human per- the transformations themselves, indicating the adjustment of
ceiver. In fact, we can draw a simple Visualization Reference these transformations by user-operated controls. It is the rapid
[AQ19] Model of these mappings (Fig. 26.10). Arrows follow from Raw reciprocal reaction between the generation of images by ma-
[AQ20] Data (data in some idiosyncratic format) on the left, though a chine and the selection and parametric adjustment of those im-
set of Data Transformations into Data Tables (canonical de- ages, giving rise to new images that gives rise to the attractive
scriptions of data in a variables x cases format extended to in- power of interactive information visualization.
clude metadata). The most important mapping is the arrow
from Data Tables to Visual Structures (structures that combine
values an available vocabulary of visual elements—spatial sub- Data Structures
strates, marks, and graphical properties). Visual Structures can
be further transformed by View Transformations, such as vi- It is convenient to express Data Tables as tables of objects and
sual distortion or 3D viewing angle, until it finally forms a View their attributes, as in Table 26.3. For example, in the FilmFinder, [AQ21]
that can be perceived by human users. Thus, Raw Data might the basic objects (or “cases”) are films. Each film is associated
start out as text represented as indexed strings or arrays. These with a number of attributes or variables, such as title, stars, year
might be transformed into document vectors, normalized vec- of release, genre type, and so forth. The vertical double black
tors in a space with dimensionality as large as the number of line in the table separates data in the table to the left of the line

FIGURE 26.10. Reference model for visualization (Card et al., 1999). Visualization can be
described as the mapping of data to visual form that supports human interaction in a workplace
for visual sense making.
 520 • CARD

TABLE 26.3. A Data Table About Films by dividing them into ranges. For example, film lengths [0, 360]
minutes (type Q) can be broken into the ranges (type O),
FilmID 230 105 540 ...
Title Goldfinger Ben Hur Ben Hur ... [0, 360] minutes → SHORT, MEDIUM, LONG.
Director Hamilton Wyler Niblo ...
Actor Connery Heston Novarro ... This common transformation is called “classing,” because it
Actress Blackman Harareet McAvoy ... maps values onto classes of values. It creates an accessible sum-
Year 1964 1959 1926 ...
mary of the data, although it loses information. In the other di-
Length 112 212 133 ...
Popularity 7.7 8.2 7.4
rection, nominal variables can be transformed to ordinal values
Rating PG G G ...
FilmType Action Action Drama ... N→O
Source: (Card et al., 1999).
based on their name. For example, film titles {GOLDFINGER, BEN
HUR, STAR WARS} can be sorted lexicographically
from the metadata, expressed as variable names, to the left of
the line. The horizontal black line across the table separates in- {GOLDFINGER, BEN HUR, STAR WARS} →
put variables from output variables—that is, the table can be BEN HUR, GOLDFINGER, STAR WARS.
thought of as a function.
Strictly speaking, we have not transformed their values, but in
f(input variables)  output variables. many uses (e.g., building alphabetically arranged dictionaries of
words or sliders in the FilmFinder), we can act as if we had.
So, Variable scale types form an important class of metadata that,
as we shall see, is important for proper information visualiza-
Year (FilmID  105)  1959. tion. We can add scale type to our Data Table in Table 26.3 to-
gether with cardinality or range of the data to give us essentially
a codebook of variables as in Table 26.5. [AQ25]
Variables imply a scale of measurement, and it is important to
keep these straight. The most important to distinguish are
Visual Structures
N  Nominal (are only  or  to other values)
O  Ordinal (obeys a  relation) Information visualization maps data relations into visual form. At
Q  Quantitative (can do arithmetic on them) first, it might seem that a hopelessly open set of visual forms can
result. Careful reflection, however, reveals what every artist
A nominal variable N is an unordered set, such as film titles knows: that visual form is subject to strong constraints. Visual
{Goldfinger, Ben Hur, Star Wars}. An ordinal variable O is a tu- form that reflects the systematic mapping of data relations onto
ple (ordered set), such as film ratings G, PG, PG-13, R. A quan- visual form, as in information visualization or data graphics, is
titative variable Q is a numeric range, such as film length [0, 360]. subject to even more constraints. It is a genuinely surprising
In addition to the three basic types of variables, subtypes fact, therefore, that most information visualization involves the
represent important properties of the world associated with mapping data relations onto only a half dozen components of
specialized visual conventions. We sometimes distinguish the visual encoding:
subtype Quantitative Spatial (Qs) for intrinsically spatial vari-
ables common in scientific visualization and the subtype Quan- 1. Spatial substrate
titative Geographical (Qg) for spatial variables that are specifi- 2. Marks
cally geophysical coordinates. Other important subtypes are 3. Connection
similarity metrics Quantitative Similarity (Qm), and the tempo- 4. Enclosure
ral variables Quantitative Time (Qt) and Ordinal Time (Ot). We 5. Retinal properties, or
can also distinguish Interval Scales (I) (like Quantitative Scales, 6. Temporal encoding
but since there is not a natural zero point, it is not meaningful to
take ratios). An example would be dates. It is meaningful to sub- Of these mappings, the most powerful is how data are
[AQ22] tract two dates ( June 5, 2002 – June 3, 2002  2 days), but it mapped onto the spatial substrate—that is, how data are
[AQ23] does not make sense to divide them ( June 5, 2002  June 23, mapped into spatial position. In fact, one might say that the de-
2002  Undefined). Finally, we can define an Unstructured sign of an information visualization consists first of deciding
Scale (U), whose only value is present or absent (e.g., an error which variables are going to get the spatial mappings, and then
[AQ24] flag). The scales are summarized in Table 26.4. how the rest of the variables are going to make do with the cod-
Scale types can be altered by transformations, and this prac- ing mappings that are left.
tice is sometimes convenient. For example, quantitative variables
can be mapped by data transformations into ordinal variables Spatial substrate. As we have just said, the most impor-
tant choice in designing an information visualization is which
Q→O variables are going to map onto spatial position. This decision
 26. Information Visualization • 521

TABLE 26.4. Classes of Data and Visual Elements

Data Classes Visual Classes

Class Description Example Description Example

U Unstructured (can only distinguish ErrorFlag Unstructured (no axis, indicated merely Dot
presence or absence) whether something is present or absent)
N Nominal (can only distinguish whether {Goldfinger, Nominal Grid (a region is divided into Colored circle
two values are equal) Ben Hur, subregions, in which something can be
Star Wars} present or absent)
O Ordinal (can distinguish whether one Small, Medium, Ordinal Grid (order of the subregions is Alpha slider
value is less or greater but not Large meaningful)
difference or ratio)
I Interval (can do subtraction on values, [10 Dec. 1978– Interval Grid (region has a metric but no Year axis
but no natural zero and can’t compute 4 Jun. 1982] distinguished origin)
ratios)
Q Quantitative (can do arithmetic on values) [0–100] kg Quantitative Grid (a region has a metric) Time slider
Qs —Spatial variables [0–20] m —Spatial grid
Qm —Similarity [0–1] —Similarity space
Qg —Geographical coord. [30°N–50°N]Lat. —Geographical cood.
Qt —Time variable [10–20] sec —Time grid

TABLE 26.5. Data Table with Meta-Data Besides these, it is convenient to make additional distinctions
Describing the Types of the Variables for frequently used subtypes, such as Spatial axes (Qs).
Axes can be linear or radial; essentially, they can involve any
FilmID N 230 105 ...
of the various coordinate systems for describing space. Axes are
Title N Goldfinger Ben Hur ... an important building block for developing Visual Structures.
Director N Hamilton Wyler ...
Based on the Data Table for the FilmFinder in Table 26.5, we rep-
Actor N Connery Heston ...
resent the scatterplot of as composed of two orthogonal quan-
Actress N Blackman Harareet ...
Year Qt 1964 1959 ... titative axes:
Length Q 112 212 ...
Popularity Q 7.7 8.2 Year → Qx,
Rating O PG G ... Popularity → Qy.
FilmType N Action Action ...
Source: (Card et al., 1999). The notation states that the Year variable is mapped to a
quantitative X-axis and the Popularity variable is mapped to a
quantitative Y-axis. Other axes are used for the FilmFinder query
widgets. For example, an ordinal axis is used in the radio but-
gives importance to spatially encoded variables at the expense tons for film ratings,
of variables encoded using other mappings. Space is perceptu-
ally dominant (MacEachren, 1995); it is good for discriminating Ratings → Oy.
values and picking out patterns. It is easier, for example, to iden-
tify the difference between a sine and a tangent curve when en- and a nominal axis is used in the radio buttons for film type,
coded as a sequence of spatial positions than as a sequence of
color hues. FilmType → Nx.
Empty space itself, as a container, can be treated as if it had
metric structure. Just as we classified variables according to their Marks. Marks are the visible things that occur in space.
scale type, we can think of the properties of space in terms of There are four elementary types of marks (Fig. 26.11): [AQ26]
the scale type of an axis of space (cf. Engelhardt, Bruin, Janssen,
& Scha, 1996). Axis scale types correspond to the variable scale 1. P  Points (0D),
types (see Table 26.4). The most important axes are 2. L  Lines (1D),
3. A  Areas (2D), and
U  Unstructured (no axis, indicated merely whether some- 4. V  Volumes (3D).
thing is present or absent)
N  Nominal Grid (a region is divided into subregions, in Area marks include surfaces in three dimensions, as well as 2D-
which something can be present or absent) bounded regions.
O  Ordinal Grid (the ordering of these subregions is mean- Unlike their mathematical counterpart, point and line marks
ingful), and actually take up space (otherwise, they would be invisible) and
Q  Quantitative Grid (a region has a metric). may have properties such as shape.
 522 • CARD

FIGURE 26.11. Types of marks. FIGURE 26.12. Retinal properties (Card et al., 1999). The six
retinal properties can be grouped by whether they form a scale
with a natural zero point (extend) and whether they deal with
spatial distance or orientation (spatial).
Connection and enclosure. Point marks and line marks
can be used to signify other sorts of topological structure: graphs
and trees. These allow showing relations among objects with-
out the geometrical constraints implicit in mapping variables
onto spatial axes. Instead, we draw explicit lines. Hierarchies and
other relationships can also be encoded using enclosure. En- to changes in mark position and the mark’s retinal properties.
closing lines can be drawn around subsets of items. Enclosure We need to distinguish between temporal data variables to be
can be used for trees, contour maps, and Venn Diagrams. visualized

Retinal properties. Other graphical properties were Qt → some visual representation

called retinal properties by Bertin (1967/1983), because the
retina of the eye is sensitive to them independent of position. and animation, that is, mapping a variable into time,
For example, the FilmFinder in Fig. 26.1 uses color to encode in-
formation in the scatterplot:
some variable → Time.

FilmID(FilmType) → P(Color)
Time as animation could encode any type of data (whether it
would be an effective encoding is another matter). Time as ani-
This notation says that the FilmType attribute for any FilmID mation, of course, can be used to visualize time as data.
case is visually mapped onto the color of a point.
[AQ27] Figure 26.12 shows Bertin’s six “retinal variables” separated
Qt→ Time.
into spatial properties and object properties according to which
area of the brain they are believed to be processed (Kosslyn,
1994). They are sorted according to whether the property is This is natural, but not always the most effective encoding. Map-
good for expressing the extent of a scale (has a natural zero ping time data into space allows comparisons between two points
point), or whether its principal use is for differentiating marks in time. For example, if we map time and a function of time into
(Bertin, 1977/1981). Spatial position, discussed earlier as basic space (e.g., time and accumulated rainfall),
visual substrate, is shown in the position it would occupy in this
classification. Qt → Qx [make time be the X-axis]
Other graphical properties have also been proposed for en- f(Qt) → Qy, [make accumulated rainfall be the Y-axis,
coding information. MacEachren (1995) has proposed (a) crisp-
ness (the inverse of the amount of distance used to blend two then we can directly experience rates as visual linear slope, and
areas or a line into an area), (b) resolution (grain with raster or we can experience changes in rates as curves. This encoding of
vector data will be displayed), (c) transparency, and (d) arrange- time into space for display allows us to make much more pre-
ment (e.g., different ways of configuring dots). He further pro- cise judgments about rates than would be possible from encod-
posed dividing color into (a) value (essentially, the gray level of ing time as time. Another use of time as animation is similar to
Fig. 26.12), (b) hue, and (c) saturation. Graphical properties the unstructured axes of space. Animation can be used to en-
from the perception literature that can support preattentive hance the ability of the user to keep track changes of view or
processing have been suggested candidates for coding variables visualization. If the user clicks on some structure, causing it to
such as curvature, lighting direction, or direction of motion (see enlarge and other structures to become smaller, animation can
Healey, Booth, and Enns, 1995). All of these suggestions require effectively convey the change and the identity of objects across
further research. the change, whereas simply viewing the two end states is con-
fusing. Another use is to enhance a visual effect. Rotating a com-
Temporal encoding. Visual Structures can also tempo- plicated object, for example, will induce 3D effects (hence, allow
rally encode information; human perception is very sensitive better reading of some visual mappings).
 26. Information Visualization • 523

Expressiveness and Effectiveness and how they are mapped onto the Visual Structure. Note that
the nominal data of the PG ratings is mapped onto a nominal
Visual mappings transform Data Tables into Visual Structure and visualization technique (colors). Note also, that names of direc-
then into a visual image. This image is not just an arbitrary im- tors and stars (nominal variables) are raised to ordinal variables
age. It is an image that has a particular meaning it must express. (through alphabetization), and then mapped onto an ordinal
That meaning is the data relation of which it is the visual trans- axis. This is, of course, a common way to handle searching
formation. We can think of the image as a sentence in a visual among a large number of nominal items.
language (Mackinlay, 1986b) that expresses the relations in the Some properties are more effective than others for encoding
Data Table. To be a good information visualization, the map- information. Position is by far the most effective all-around rep-
pings must satisfy some constraints. The first constraint is that resentation. Many properties are more effective for some types
the mapping must be expressive. A visualization is said to be of data than for others. Table 26.7 gives an approximate evalua-
expressive if and only if it encodes all the data relations intended tion for the relative effectiveness of some encoding techniques
and no other data relations. The first part of expressiveness based on (MacEachren, 1995). We note that spatial position is ef-
turns out to be easier than the second. Suppose we plot Film- fective for all scale types of data. Shape, on the other hand, is
[AQ28] Type against Year using the data-to-visual mapping in Fig. 26.13. only effective for nominal data. Gray scale is most effective for
The problem of this mapping is that the nominal movie rating ordinal data. Such a chart can suggest representations to a vi-
data are expressed by a quantitative axis. That is, we have tried sualization designer.
to map

FilmType(N) → Position(Q).
Taxonomy of Information Visualizations
In so doing, we have visually expressed all the data relation,
We have shown that the properties of data and visual repre-
but the visualization also implies relationships that do not ex-
sentation generally constrain the set of mappings that form the
ist. For example, the 1959 version of Ben Hur does not have
basis for information visualizations. Taken together, these con-
a film type that is five times greater than the 1926 version of
straints form the basis of a taxonomy of information visualiza- [AQ32]
Ben Hur, as implied in the figure. Wisely, the authors of the
tions. Such a taxonomy is given in Table 26.8. Visualizations are
FilmFinder chose the mapping
grouped into four categories. First are Simple Visual Struc-
tures, the static mapping of data onto multiple spatial dimen-
FilmType(N ) → Color(N ). sions, trees, or networks plus retinal variables, depicted in Fig.
26.10. Here it is worth distinguishing two cases. There is a per-
Of course, there are circumstances in which color could be ceptual barrier at three (or, in special cases, four) variables, a
read as ordinal, or even possibly quantitative, but the miscella- limit of the amount of data that can be perceived as an imme-
neous order of the buttons in Fig. 26.1 discourages such an in- diate whole. Bertin (1977, 1981) called this elementary unit of
terpretation and the relatively low effectiveness of color for this visual data perception the “image”. Although this limit has not
[AQ29] purpose in Table 26.7 also discourages this interpretation. been definitively established in information visualization by em-
[AQ30] Table 26.6 shows the mappings chosen by authors of the pirical research, there must be a limit somewhere or else peo-
[AQ31] FilmFinder. The figure shows the Data Table’s metadata and data ple could simultaneously comprehend a thousand variables.
We therefore divide visualizations into those that can be com-
prehended in an elementary perceptual grasp (three, or in spe-
cial cases, four variables)—let us call these direct reading
visualizations—and those more complex than that barrier—
which we call articulated reading visualizations, in which
multiple actions are required.
Beyond the perceptual barrier, direct composition of data re-
lationships in terms of 1, 2, or 3 spatial dimensions plus re-
maining retinal variables is still possible, but rapidly diminishes
in effectiveness. In fact, the main problem of information visual-
ization as a discipline can be seen as devising techniques for ac-
celerating the comprehension of these more complex n-variable
data relations. Several classes of techniques for n-variable visu-
alization, which we call Composed Visual Structures, are based
on composing Simple Visual Structures together by reusing
their spatial axes. A third class of Visual Structures—Interac-
tive Visual Structures—comes from using the rapid interaction
capabilities of the computer. These visualizations invoke the
parameter-controlling arrows of Fig. 26.10. Finally, a fourth class
FIGURE 26.13. Mapping from data to visual form that violates of visualizations—Attention-Reactive Visual Structures—comes
expressiveness criterion. from interactive displays where the system reacts to user actions
524 • CARD

TABLE 26.6. Meta-Data and Mappings of Data onto Visual Structure in the FilmFinder

Source: (Card et al., 1999).

TABLE 26.7. Relative Effectiveness of Position and Retinal Encodings

Spatial Q O N Object Q O N
Extent (Position)    Gray Scale   
Size    Color   
Differential Orientation    Texture   
Shape   
Source: (Card et al., 1999).

by changing the display, even anticipating new displays, to lower cases), while another was used to encode the objects’ values. Ex-
to cost of information access and sensemaking to the user. To amples of this notation appear in Table 26.8 and Fig. 26.21.
summarize,

I. Simple Visual Structures SIMPLE VISUAL STRUCTURES

Direct Reading
Articulated Reading The design of information visualizations begins with mappings
II. Composed Visual Structures from variables of the Data Table into the Visual Structure. The
Single-Axis Composition basic strategy for the visualization designer could be described
Double-Axis Composition as follows:
Recursive Composition
III. Interactive Visual Structure 1. Determine which variables of the Analytic Abstraction to
IV. Attention-Reactive Visual Structure map into spatial position in the Visual Structure.
2. Combine these mappings to increase dimensionality (e.g.,
These classes of techniques may be combined to produce vi- by folding).
sualizations that are more complex. To help us keep track of the 3. Use retinal variables as an overlay to add more dimensions.
variable mapping into visual structure, we will use a simple short- 4. Add controls for interaction.
hand notation for listing the element of the Visual Structure that 5. Consider attention-reactive features to expand space and
the Data Table has mapped into. We will write, for example, manage attention.
[XYR2] to note that variables map onto the X-axis, the Y-axis, and
two retinal encodings. [OX] will indicate that the variables map We start by considering some of the ways in which variables can
onto one spatial axis used to arrange the objects (that is, the be mapped into space.
 26. Information Visualization • 525

TABLE 26.8. Taxonomy of Information Visualization Techniques

1-Variable More generally, many single-variable visualizations are in the

form v  f(o), where v is a variable attribute and o is the object.
Figure 26.14(c) is of this form and uses the Y-axis to encode the
One-variable visual displays may actually use more than one vi- variable and the X-axis for the objects. Note that if the objects
sual dimension. This is because the data variable or attribute is are, as usual, nominal, then they are reorderable: sorting the ob-
displayed against some set of objects using some mark and be- jects on the variable produces easily perceivable visual patterns.
cause the mark itself takes space. Or, more subtly, it may be be- For convenience, we have used rectangular coordinates, but any
cause one of the dimensions is used for arranging the objects other orthogonal coordinates could be used as the basis of de-
and another for encoding via position the variable. A simple ex- composing space. Figure 26.14(d) uses  from polar coordinates
ample would be when the data are just visually mapped into a to encode, say, percentage voting for different presidential can-
[AQ33] simple text list as in Fig. 26.14(a). The objects form a sequence didates. In Fig. 26.14(e), a transformation on the data side has
on the Y-dimension, and the width of the marks (the text de- transformed variable o into a variable representing the distribu-
scriptor) takes space in the X-dimension. By contrast, a one-di- tion, then mapped that onto points on the Y-axis. In Fig.
mensional scattergraph (Fig. 26.14[b]) does not use a dimen- 26.14(f ), another transformation on the data side has mapped
sion for the objects. Here, the Y-axis is used to display the this distribution into 2nd quartiles, 3rd quartiles, and outlier
attribute variable (suppose these are distances from home of points, which is then mapped on the visual side into a box plot
gas stations); the objects are encoded in the mark (which takes on the Y-axis. Simple as they are, these techniques can be very
a little bit of the X-dimension). useful, especially in combination with other techniques.
 526 • CARD

of Fig. 26.15(b) has a one-variable visualization that serves as a

legend for the main visualization: it associates color hues with
dates and allows the selection of date ranges.

2-Variables

As we increase the number of variables, it is apparent that their

mappings form a combinatorial design space. Figure 26.16 [AQ35]
schematically plots the structure of this space, leaving out the
use of multiple lower variable diagrams to plot higher variable
combinations. Two-variable visualizations can be thought of
as a composition of two elementary axes (Bertin, 1977, 1981;
Mackinlay, 1986b), which use a single mark to encode the posi-
tion on both those axes. Mackinlay called this mark composi-
tion, and it results in a 2D scattergraph (Fig. 26.16[g]). Note that
instead of mapping onto two positional visual encodings, one
positional axis could be used for the objects, and the data vari-
ables could be mapped onto a position encoding and a retinal
encoding (size), as in Fig. 26.16(f ).

3-Variables and Information Landscapes

By the time we get to three data variables, a visualization ca be

produced in several ways. We can use three separate visual di-
mensions to encode the three data variables in a 3D scatter-
graph (Fig. 26.16[j]). We could also use two spatial dimensions
and one retinal variable in a 2D retinal scattergraph (Fig.
26.16[k]). Or we could use one spatial dimension as an object
dimension, one as a data attribute dimension, and one two reti-
FIGURE 26.14. 1-variable visual abstractions. nal encodings for the other variables, as in an object chart such
as in Fig 26.16(i). Because Fig. 26.16(i) uses multiple retinal en-
codings, however, it may not be as effective as other techniques.
Notice that because they all encode three data variables, we
One special, but common, problem is how to visualize very have classified 2D and 3D displays together. In fact, one popular
large dimensions. This problem occurs for single-variable visu- 3-variable information visualization that lies between 2D and 3D
alizations, but may also occur for one dimension of a multi-vari- is the information landscape (Fig. 26.16[m]). This is essentially
[AQ34] able visualization. Figure 26.15 shows several techniques for a 2D scattergraph with one data variable extruded into the third
handing the problem. In Fig. 26.15(a) (Freeman & Fertig, 1995), spatial dimension. Its essence is that two of the spatial dimen-
the visual dimension is laid out in perspective. Even though sions are more tightly coupled and often relate to a 2D visual-
each object may take only one or a few pixels on the axis, the ization. For example, the two dimensions might form a map
objects are actually fairly large and selectable in the diagram. In with the bars showing the GDP of each region.
Fig. 26.15(b) (Eick, Steffen, & Sumner, 1992), the objects (rep- Another special type of 3-variable information visualization is
resenting lines of code) are laid out on a folded Y-axis. When the a 2D information topography. In an information typography,
Y-axis reaches the bottom of the page, it continues offset at the space is partly defined by reference to external structure. For ex-
top. In Fig. 26.15(c) (Keim & Kriegel, 1994), the axis is wrapped ample, the topography of Fig. 26.17(a) is a map of San Francisco, [AQ36]
in a square spiral. Each object is a single pixel, and its value is requiring two spatial variables. The size of blue dots indexes
coded as the retinal variable color hue. The objects have been the number of domain names registered to San Francisco street
sorted on another variable; hence, the rings show the correla- addresses. Looking at the patterns in the visualization shows
tion of this attribute with that of the sorting attribute. that Internet addresses have especially concentrated in the Mis-
One-variable visualizations are also good parts of controls. sion and South of Mission districts. Figure 26.17(a) uses a topog-
Controls, in the form of slides, also consume considerable space raphy derived from real geographical space. Various techniques,
on the display (for example, the controls in Fig. 26.1) that could such as multidimensional scaling, factor analysis, or connec-
be used for additional information communication. Figure tionist self-organizing algorithms, can create abstract spaces
26.15(d) shows a slider on whose surface is a distribution rep- based on the similarities among collections of documents or
resentation of the number of objects for each value of the in- other objects. These abstract similarity spaces can function like
put variable, thereby communicating information about the a topography. An example can be seen in Fig. 26.17(b), where
slider’s sensitivity in different data ranges. The slider on the left the pages in a Web site are depicted as regions in a similarity
 26. Information Visualization • 527

FIGURE 26.15. Uses of 1-variable visual abstractions.

[AQ37] space. To create this diagram2, a web crawler crawls the site and ables, and the possibilities for four variables are shown in Fig.
[AQ38]
indexes all the words and pages on the site. Each page is then 26.16. These diagrams can be understood, but at the cost of pro-
turned into a document vector to represent the semantic con- gressively more effort as the number of variables increases. It
tent of that page. The regions are created using a neural net- would be very difficult to understand an [XYR20] retinal scatter-
work learning algorithm (see Lin, Soergel, & Marchionini (1991)). graph, for example.
This algorithm organizes the set of Web pages into regions. A
visualization algorithm then draws boundaries around the re-
gions, colors them, and names them. The result, called a Kahonen Trees
diagram after its original inventor, is a type of retinal similarity
topography. An interesting alternative to showing variable values by spatial
Information landscapes can also use marks that are surfaces. positioning is to use explicitly drawn linkages of some kind.
[AQ39] In Fig. 26.18(a), topics are clustered on a similarity surface, and Trees are the simplest form of these. Trees map cases into sub-
the strength of each topic is indicated by a 3D contour. A more cases. One of the data variables in a Data Table (for example, the
extreme case is Fig. 26.18(b), where an information landscape is variable ReportsTo in an organization chart) is used to define
established in spherical coordinates, and the amount of ozone the tree. There are two basic methods for visualizing a tree: (a)
is plotted as a semitransparent overlay on the -axis. Connection and (b) Enclosures.

Connection. Connection uses lines to connect marks sig-

n-Variables nifying the nodes of the tree. Logically, a tree could be drawn
merely by drawing lines between objects located randomly po-
Beyond three variables, direct extensions of the methods we sitioned on the plane, but such a tree would be visually unread-
have discussed become less effective. It is possible, of course able. Positioning in space is important. Figure 26.20(a) is a tree [AQ40]
to make plots using two spatial variables and n–2 retinal vari- from Charles Darwin’s notebook (Robin, 1992) drawn to help

2
This figure is produced by a program called SiteMap by Xia Lin and associates. See http://faculty.cis.drexel.edu/sitemap/index.html. [AQ56]
FIGURE 26.16. Simple Visual Structures..

528
 26. Information Visualization • 529

FIGURE 26.18. 3D information surface topographies.

tree structure and space makes trees relatively easy to lay out
and interpret, compared to generalized networks. Hierarchical
displays are important not only because many interesting col-
lections of information, such as organization charts or tax-
FIGURE 26.17. Retinal information topographies. onomies, are hierarchical data, but also because important col-
lections of information, such as Web sites, are approximately
hierarchical. Whereas practical methods exist for displaying
him work out the theory of evolution. Lines proceed from an- trees up to several thousand nodes, no good methods exist for
cestor species to new species. Note that even in this informal displaying general graphs of this size. If a visualization problem
setting intended for personal use that the tree uses space sys- involves the displaying of network data, a practical design
tematically (and opportunistically). There are no crossed lines. heuristic is to see whether the data might not be forced into a
A common way of laying out trees is to have the depth in the display as a modified tree, such as a tree with a few non-tree
tree map onto one ordinal access as in Fig. 26.20(b), while the links. A significant disadvantage of trees is that as they get large,
other axis is nominal and used to separate nodes. Of course, they acquire an extreme aspect ratio, because the nodes expand
trees could also be mapped into other coordinate systems: for exponentially as a function with depth. Consequently, any suffi-
example, there can be circular trees in which the r-axis repre- ciently large tree (say, 1000 nodes) resembles a straight line.
sents depth and the -axis is used to separate nodes as in the Circular trees such as Fig. 26.20(c) are one way of trying to buy
3 more space to mitigate this problem. Another disadvantage of
[AQ41] representation of the evolution species in Fig. 26.20(c). It is be-
cause trees have no cycles that one of the spatial dimensions trees is the significant empty space between nodes to make
can be used to encode tree depth. This partial correlation of their organization easily readable. Various tricks can be used to

3
[AQ57] This figure is from David Hillis, University of Texas.
530 • CARD

wrap parts of the tree into this empty space, but at the expense
of the tree’s virtues of readability.

Enclosure. Enclosure uses lines to hierarchically enclose

nested subsets of the tree. Figure 26.20(d) is an enclosure tree
encoding of Darwin’s tree in Fig. 26.20(a). We have already seen
one attempt to use tree enclosure, TreeMaps (Fig. 26.5).
TreeMaps make use of all the space and stays within prescribed
space boundaries, but they do not represent the nonterminal
nodes of the tree very well and similar leaves can have wildly dif-
ferent aspect ratios. Recent variations on TreeMaps found ways
to “squarify” nodes (Shneiderman & Wattenberg, 2001), miti-
gating this problem.

Networks

Networks are more general than trees and may contain cycles.
Networks may have directional links. They are useful for de-
scribing communication relationships among people, traffic in a
telephone network, and the organization of the Internet. Con-
tainment is difficult to use as a visual encoding for network re-
lationships, so most networks are laid out as node and link dia-
grams. Unfortunately, straightforward layouts of large node and
link diagrams tend to resemble a large wad of tangled string.
We can distinguish the same types of nodes and links in net-
work Visual Structures that we did for spatial axes: (a) Unstruc-
FIGURE 26.19. Trees. tured (unlabeled), (b) Nominal (labeled), (c) Ordinal (labeled

FIGURE 26.20. Network methods.

 26. Information Visualization • 531

with an ordinal quantity), or (d) Quantitative (weighted links). near each other on the array. Suspicious patterns are visible be-
Retinal properties, such as size or color, can be used to encode cause of the sparseness of the network.
information about links and nodes. As in the case of trees, spa- The insightful display of large networks is difficult enough
tial positioning of the nodes is extremely important. Network vi- that many information visualization techniques depend on in-
sualizations escape from the strong spatial constraints of sim- teractivity. One important technique, for example, is node ag-
ple Visual Structures only to encounter another set of strong gregation. Nodes can be aggregated to reduce the number of
spatial constraints of node links crossing and routing. Networks links that have to be drawn on the screen. Which nodes are ag-
and trees are not so much an alternative of direct of the direct gregated can depend on the portion of the network on which
graphical mappings we have discussed so far as they are another the user is drilling down. Similarly, the sets of nodes can be in-
set of techniques that can be overlaid on these mappings. Small teractively restricted (e.g., telephone calls greater than a cer-
node and link diagrams can be laid out opportunistically by tain volume) to reduce the visualization problem to one within
hand or by using graph drawing algorithms that have been de- the capability of current techniques.
veloped (Battista, Eades, Tamassia, & Tollis, 1994; Cruz & Tamas-
sia, 1998; Tamassia, 1996) to optimize minimal link crossing,
symmetry, and other aesthetic principles.
For very large node and link diagrams, additional organizing COMPOSED VISUAL STRUCTURES
principles are needed. If there is an external topographic struc-
ture, it is sometimes possible to use the spatial variables associ- So far, we have discussed simple mappings from data into spa-
ated with the nodes. Figure 26.20(a) shows a network based on tial position axes, connections and enclosures, and retinal vari-
call traffic between cities in the United States (Becker, Eick, & ables. These methods begin to run into a barrier around three
Wilks, 1995). The geographical location of the cities is used to variables as the spatial dimensions are used up and as multiple
lay out the nodes of the network. Another way to position of the less efficient retinal variables needed. Most interesting
nodes is by associating nodes with positions in a similarity problems involve many variables. We shall therefore look at a
space, such the nodes that have the strongest linkages to each class of methods that reuse precious spatial axes to encode vari-
other are closest together. There are several methods for com- ables. This is done by composing a compound Visual Structure
puting node nearness in this way. One is to use multidimen- out of several simple Visual Structures. We will consider five sub-
sional scaling (MDS) (Fairchild, Poltrock, & Furnas, 1988). An- classes of such composition: (a) mark composition, (b) case
other is to use a “spring” technique, in which each link is composition, (c) single-axis composition, (d) double-axis com-
associated with a Hooke’s Law spring weighted by strength of position, and (e) recursive composition. Schematically, we il-
[AQ43]
association and the system of springs is solved to obtain node lustrate these possibilities in Fig. 26.21.
position. Eick and Willis (1993) have argued that the MDS tech-
nique places too much emphasis on smaller links. They have de- Single-axis composition. In single-axis composition,
[AQ42] rived an alternative that gives clumpier (and hence, more visu- multiple variables that share a single axis are aligned using that
ally structured) clusters of nodes. If positioning of nodes axis, as illustrated in Fig. 26.21(a). An example of single-axis
corresponds perfectly with linkage information, then the links composition is a method due to Bertin called permutation
do not add more visual information. If positioning does not cor- matrices (Bertin, 1977/1981). In a permutation matrix (Fig.
respond at all with linkage information, then the diagram is ran- 26.16[o], for example), one of the spatial axes is used to repre-
dom and obscure. In large graphs, node positions must have a sent the cases and the other a series of bar charts (or rows of cir-
partially correlated relationship to linkage in order to allow the cles of different size or some other depiction of the value of
emergence of visual structure. Note that this is what happens each variable) to represent the values. In addition, bars for val-
in the telephone traffic diagram Fig. 26.20(a). Cities are posi- ues below average may be given a different color, as in Fig. 26.7,
tioned by geographical location. Communication might be ex- in order to enhance the visual patterns. The order of the objects
pected to be higher among closer cities, so the fact that com- and the order of the variables may both be permuted until pat-
munications is heavy between coasts stands out. terns come into play. Permutation matrices were used in our ho-
A major problem in a network such as Fig. 26.20(a) is that tel analysis example. They give up direct reading of the data
links may obscure the structure of the graph. One solution is space in order to handle a larger number of variables. Of course,
to route the links so that they do not obscure each other. The as the number of variables (or objects) increases, manipulation
links could even be drawn outside the plane in the third di- of the matrices becomes more time-consuming and visual in-
mension; however, there are limits to the effectiveness of this terpretation more complex. Still, permutation matrices or their
technique. Another solution is to use thresholding, as in Fig. variants are one of the most practical ways of representing
26.20(b). Only those links representing traffic greater than a cer- multi-variable data.
tain threshold are included; the others are elided allowing us If we superimpose the bar charts of the permutation matrix
to see the most important structure. Another technique is line atop one another, and then replace the bar chart with a line link-
shortening, as in Fig. 26.20(c). Only the portion of the line near ing together the tops of the bars, we get another method for
the nodes is drawn. At the cost of giving up the precise linkage, handling multiple variables by single-axis composition—paral-
it is possible to read the density of linkages for the different lel coordinates (Inselberg, 1997; Inselberg & Dimsdale, 1990),
nodes. Figure 26.20(d) is a technique used to find patterns in an as shown in Fig. 26.22. A problem is analyzed in parallel coordi- [AQ44]
extremely large network. Telephone subscribers are repre- nates by interactively restricting the objects displayed (the lines)
sented as nodes on a hexagonal array. Frequent pairs are located in order to look at cases with common characteristics. In Fig.
532 • CARD

FIGURE 26.21. Composition types.

26.22 , parallel coordinates are used to analyze the problem of

yield from a certain processor chip. X1 is chip yield, X2 is qual-
ity, X3 through X12 are defects, and the rest of the variables are
physical parameters. The analysis, looking at those subsets of
data with high yield and noticing the distribution of lines on
the other parameters, was able to solve a significant problem in
chip processing.
Both permutation matrices and parallel coordinates allow
analyses in multi-dimensional space, because they are efficient
in the use (and reuse) of spatial position and the plane. Actually,
they also derive part of their power from being interactive. In
the case of permutation matrices, interactivity comes in re-
ordering the matrices. In the case of parallel coordinates, inter-
activity comes in selecting subsets of cases to display.

Double-axis composition. In double-axis composition,

two visual axes must be in correspondence, in which case the
cases are plotted on the same axes as a multivariable graph (Fig.
26.21[b]). Care must be taken that the variables are plotted on a
comparable scale. For this reason, the separate scales of the vari-
ables are often transformed to a common proportion change
FIGURE 26.22. Single-axis composition: parallel coordinates. scale. An example would be change in price for various stocks.
 26. Information Visualization • 533

The cases would be the years, and the variables would be the in each region (Fig. 26.21[e]). We use the term somewhat
different stocks. loosely, since regions have different types of subvisualizations.
The FilmFinder in Fig. 26.1 is a good example of a recursive vi-
Mark composition and case composition. Composi- sualization. The screen breaks down into a series of simple Vi-
tion can also fuse diagrams. We discussed that each dimension sual Structures and controls: (a) a 3-variable retinal scattergraph
of visual space can be said to have properties as summarized in (Year, Rating, FilmType)  (b) a 1-variable slider (Title)  (c) a
Table 26.4. The visual space of a diagram is composed from the 1-variable slider (Actors)  (d) a 1-variable slider (Actresses) 
properties of its axis. In mark composition (Fig. 26.21[c]), the (e) a 1-variable slider (Director)  (f ) a 1-variable slider (Film-
mark on one axis can fuse with the corresponding mark on an- Length)  (g) a 1-variable radio button control (Rating) 
other axis to form a single mark in the space formed by the two (h) a 1-variable button-set (FilmType).
axes. Similarly, two object charts can be fused into a single dia- Three types of recursive composition deserve special men-
gram by having a single mark for each case. We call this latter tion: (a)2D-in-2D, (b) marks-in-2D, and (c) 3D-in-3D. An exam-
form case composition Fig. 26.21(d). ple of 2D-in-2D composition is the “prosection matrix”
(Tweedie, Spence, Dawkes, & Su, 1996) shown in Fig. 26.23(a). [AQ45]
Recursive composition. Recursive composition divides Each smaller square in the prosection matrix represents a pair of
the plane (or 3D space) into regions, placing a subvisualization parameters plotted against each other. The coloring shows

FIGURE 26.23. Recursive composition.

534 • CARD

which values of the plotted pair give excellent (red region) or the visualization. If the user clicks on a button or moves a slider,
partly good (gray regions) performance for the design of some the system needs to update the display in less than 0.1 sec. Ani-
device. The arrangement of the individual matrices into a su- mation frames need to take less than 0.1 sec. The second time
permatrix redefines the spatial dimensions (that is, associates constant, 1 sec, is the time to complete an immediate action, for
it with different variables) within each of the cells, and the cells example, an animated sequence such as zooming in to the data
themselves are arranged in an overall scheme that systematically or rotating a tree branch. The third time constant 10 sec (mean-
uses space. In this way, the precious spatial dimension is effec- ing somewhere in the 5 to 30 sec interval) is the time for com-
tively expanded to where all the variables can reuse it. An im- pleting some cognitive action, for example deleting an element
portant property of techniques similar to this one is that space from the display. Let us consider a few well-known techniques for
is defined at more than one grain size, and these levels of grain interactive information visualizations.
become the basis for a macro-micro reading.
An example of marks-in-2D composition in the use of “stick Dynamic queries. A general paradigm for visualization
figure” displays. This is an unusual type of visualization in which interaction is dynamic queries, the interaction technique used
the recursion is within the mark instead of within the use of by the FilmFinder in Fig. 26.1. The user has a visualization of the
space. Figure 26.23(b) shows a mark that is itself composed of data and a set of controls, such as sliders, by which subsets of
submarks. The mark is a line segment with four smaller line seg- the Data Table can be selected. For example, Table 26.9 shows [AQ46]
ments protruding from the ends. Four variables are mapped the mappings of the Data Table and controls for the FilmFinder.
onto angle of these smaller line segments and a fifth onto the The sliders and other controls will select which subset of the
angle of the main line segment. Two additional variables are data is going to be displayed. In the FilmFinder, the control for
mapped onto the position of this mark in a 2D display. A typi- Length is a two-sided slider. Setting one end to 90 minutes and
cal result is the visualization in Fig. 26.23(c), which shows five the other end to 120 minutes will select for display only those
weather variables around Lake Ontario, the outline of which cases of the Data Table whose year variable lies between these
clearly appears in the figure. limits. The display needs to change within the 0.1 sec of chang-
Feiner and Beshers (1990) provided an example of the third ing the slider.
recursive composition technique, 3D-in-3D composition. Sup-
pose a dependent variable is a function of six continuous vari- Magic lens (movable filter). Dynamic queries is one
ables, y  f(x, y, z, w, r, s). Three of these variables are mapped type of interactive filter. Another type is a movable filter that can
onto a 3D coordinate system. A position is chosen in that space, be moved across the display, as in Fig. 26.24(a). These magic [AQ47]
say, x1, y1, z1. At that position, a new 3D coordinate system is lenses are useful when it is desired to filter only some of the
presented with a surface defined by the other three variables display. For example, a magic lens could be used with a map that
(Fig. 26.23[d]). The user can thus view y  f(x1, y1, z1, w, r, s). showed the population of any city it was moved over. Multiple
The user can slide the second-order coordinate system to any magic lenses can be used to cascade filters.
location in the first, causing the surface to change appropriately.
Note that this technique combines a composed visual inter- Overview  detail. We can think of an overview  detail
action with interactivity on the composition. Multiple second- display (Fig. 26.24[b]) as a particular type of magic lens, one that
order coordinate systems can be displayed at the space simul- magnifies the display and has the magnified region off to the
taneously, as long as they do not overlap by much. side so as not to occlude the region. Displays have information
at different grain sizes. A GIS map may have information at the
level of a continent as well as at the level of a city. If the shape
of the continent can be seen, the display is too coarse to see
INTERACTIVE VISUAL STRUCTURES the roadways of a city. Overview  detail displays show that data
at more than one level, but they also show where the finer grain
In the examples we have considered so far, we have often seen display fits into the larger grain display. In Fig. 26.24(b), from
that information visualization techniques were enhanced by be- SeeSoft (Eick et al., 1992), a system for visualizing large software
ing interactive. Interactivity is what makes visualization a new systems, the amount of magnification in the detail view is large
medium, separating it from generations of excellent work on enough that two concatenated overview  detail displays are re-
scientific diagrams and data graphics. Interactivity means con- quired. Overview  detail displays are thus very helpful for data
trolling the parameters in the visualization reference model (Fig. navigation. Their main disadvantage is that the require coordi-
26.10). This naturally means that there are different types of in- nation of two visual domains.
teractivity, because the user could control the parameters to data
transformations, to visual mappings, or to view transformations. Linking and brushing. Overview  detail is an exam-
It also means that there are different forms of interactivity based ple of coordinating dual representations of the same data.
on the response cycle of the interaction. As an approximation, These can be coordinated interactively with linking and brush-
we can think of there being three time constants that govern in- ing. Suppose, for example, we wish to show power consump-
teractivity, which we take to be 0.1 sec, 1 sec, and 10 sec (Card, tion on an airplane, both in terms of the physical representation
Moran, & Newell, 1986) (although the ideal value of these may of the airplane and a logical circuit diagram. The two views
be somewhat less, say, 0.07 sec, 0.7 sec, and 7 sec). The first time could be shown and linked by using the same color for the
constant is the time in which a system response must be made, if same component types. Interactivity itself can be used for a dy-
the user is to feel that there is a direct physical manipulation of namic form of linking called brushing. In brushing, running the
TABLE 26.9. Visual Marks and Controls for FilmFinder

FIGURE 26.24. Interaction techniques.

535
536 • CARD

cursor over a part of one of the views causes highlighting both Focuscontext views are based on several premises: First,
in that view and in the other view. the user needs both overview (context) and detail information
(focus) during information access, and providing these in sepa-
Extraction and comparison. We can also use interac- rate screens or separate displays is likely to cost more in user
tion to extract a subset of the data to compare with another sub- time. Second, information needed in the overview may be dif-
set. An example of this is in the SDM system (Chuah, Roth, Mat- ferent from that needed in the detail. The information of the
tis, & Kolojejchick, 1995) in Fig. 26.24(c). The data are displayed overview needs to provide enough information the user to de-
in a 3D information landscape, but the perspective interferes cide where to examine next or to give a context to the detailed
with the ability to compare it. Information is therefore extracted information rather than the detailed information itself. As Fur-
from the display (leaving ghosts behind) and placed in an or- nas (1981) has argued, the user’s interest in detail seems to fall
thogonal viewing position where it can be compared using 2D. away in a systematic way with distance as information objects
It could also be dropped into another display. Interactivity become farther from current interest. Third, these two types of
makes possible these manipulations, while keeping them coor- information can be combined within a single dynamic display,
dinated with the original representations. much as human vision uses a two-level focus and context strat-
egy. Information broken into multiple displays (separate leg-
Attribute explorer. Several of these interactive tech- ends for a graph, for example) seem to degrade performance
niques are combined in the Attribute Explorer (Tweedie et al., due to reasons of visual search and working memory.
1996). Figure 26.24(d) shows information on four attributes of Furnas (1981) was the first to articulate these ideas system-
houses. Each attribute is displayed by a histogram, where each atically in his theory of fisheye views. The essence of focuscon-
square making up the histogram represents an individual text displays is that the average cost of accessing information is
house. The user selects a range of some attribute, say price. reduced by placing the most likely needed information for nav-
Those pixels making up the histogram on price have their cor- igation and detail where it is fastest to access. This can be ac-
responding pixels linked representing houses highlighted on complished by working on either the data side or the visual side
the other attributes. Those houses meeting all the criteria are of the visual reference model, Fig. 26.10. We now consider these
highlighted in one color; those houses meeting, say, all but one techniques in more detail.
are highlighted in another color. In this way, the user can tell
about the “near misses.” If the users were to relax one of the cri-
teria only a little (say, reducing price by $100), then the user
might be able to gain more on another criterion (say, reducing
Data-Based Methods
a commute by 20 miles).
Filtering. On the data side, focuscontext effects can be
achieved by filtering out which items from the Data Table are ac-
tually displayed on the screen. Suppose we have a tree of cate-
gories taken from Roget’s Thesaurus, and we are interacting
FOCUS  CONTEXT ATTENTION-REACTIVE with one of these, “Hardness.”
ABSTRACTIONS
Matter
So far, we have considered visualizations that are static map- ORGANIC
pings from Data Table to Visual Structure and those where the Vitality
mappings Data Table to Visual Structure are interactively con- Vitality in general
trolled by the user. We now consider visualizations in which the Specific vitality
machine is no longer passive, but its mappings from Visual Sensation
Structure to View are altered by the computer according to the Sensation in general
its model of the user’s degree of interest. We can, in principle, Specific sensation
associate a cost of access with every element in the Data Table. INORGANIC
Take the FilmFinder in Figure 26.3. Details about the movie Solid
“Murder on the Orient Express” are accessible at low cost in Hardness
terms of time because they are presently visible on the screen. Softness
Details of “Goldfinger,” a movie with only a mark on the dis- Fluid
play, take more time to find. Details of “Last Year at Marienbad,” Fluids in general
a movie with no mark on the display, would take much more Specific fluids
time. The idea is that with a model for predicting users’ changes
in interest, the system can adjust its displays to make costs lower Of course, this is a small example for illustration. A tree rep-
for information access. For example, if the user wants some de- resenting a program listing or a computer directory or a taxon-
tail about a movie, such as the director, the system can antici- omy could easily have thousands of lines, a number that would
pate that the user is more likely to want other details about the vastly exceed what could fit on the display and hence would
movie as well and therefore display them all at the same time: have a high cost of accessing. We calculate a degree-of-interest
The user does not have execute a separate command; the cost (DOI) for each item of the tree, given that the focus is on the
is therefore reduced. node Hardness. To do this, we split the DOI into an intrinsic
 26. Information Visualization • 537

part and a part that varies with distance from the current cen- behavior in a presidential election, voters could be broken
ter of interest and use a formula from Furnas (1981). down by sex, precinct, income, and party affiliation. As the user
drills down on, say, male Democrats earning between $25,000
DOI  Intrinsic DOI  Distance DOI and $50,000, other categories could be aggregated, providing
screen space and contextual reference for the categories of im-
[AQ48] Figure 26.25 shows schematically how to perform this com- mediate interest.
putation for our example. We assume that the intrinsic DOI of
a node is just its distance of the root (Fig. 26.25[a]). The dis-
tance part of the DOI is just the traversal distance to a node
View-Based Methods
from the current focus node (Fig. 26.25[b]; it turns out to be
convenient to use negative numbers for this computation, so Micro-macro readings. Micro-macro readings are dia-
that the maximum amount of interest is bounded, but not the grams in which “detail cumulates into larger coherent struc-
minimum amount of interest). We add these two numbers to- tures” (Tufte, 1990). The diagram can be graphically read at the
gether (Fig. 26.25 [c]) to get the DOI of each node in the tree. level of larger contextual structure or at the detail level. An ex-
Then we apply a minimum threshold of interest (5 in this ample is Fig. 26.26. The micro reading of this diagram shows [AQ49]
case) and only show nodes more interesting than that thresh- three million observations of the sleep (lines), wake (spaces),
old. The result is the reduced tree: and feeding (dots) activity of a newborn infant. Each day’s ac-
tivity is repeated three times on a line to make the cyclical as-
pect of the activity more clearly visible. The macro reading of
Matter
the diagram, emphasized the thick lines, shows the infant tran-
INORGANIC
sitioning from the natural human 25-hour cycle at birth to the
ORGANIC
24-hour solar day. The macro reading serves as context and in-
Solid
dex into the micro reading.
Hardness
Softness
Highlighting. Highlighting is a special form of micro-
Fluid
macro reading in which focal items are made visually distinc-
tive in some way. The overall set of items provides a context for
The reduced tree gives local context around the focus node and the changing focal elements.
progressively less detail farther away. But it does seem to give
the important context. Visual transfer functions. We can also warp the view
with viewing transformations. An example is a visualization
Selective aggregation. Another focuscontext tech- called the bifocal lens (Spence & Apperley, 1982). Fig. 26.27(a) [AQ50]
nique from the data side is selective aggregation. Selective ag- shows a set of documents the user would like to view, but which
gregation creates new cases in the Data Table that are aggre- is too large to fit on the screen. In a bifocal lens, documents not
gates of other cases. For example, in a visualization of voting in a central focal region are compressed down to a smaller size.
This could be a strict visual compression. It could also involve a
change in representation. We can talk about the visual com-
pression in terms of a visual transfer function Fig. 26.27(b),
sometimes conveniently represented in terms of its first deriva-
tive in Fig. 26.27(c). This function shows how many units of an

FIGURE 26.25. Degree-of-Interest calculation for fish-eye visu- FIGURE 26.26. Micro-macro reading. (Winfree, 1987). Courtesy
alization. Scientific American Library.
 538 • CARD

AU:
Please
check
art.

FIGURE 26.27. Bifocal + transfer function.

axis in the original display are mapped into how many units in Distorted views must be designed carefully so as not to dam-
the resultant display. The result could be compression or en- age important visual relationships. Bubble distortions of maps
largement of a section of the display. As a result of applying this may change whether roads appear parallel to each other. How-
visual transfer function to Fig. 26.27(a), the display is com- ever, distorted views can be designed with “flat” and “transition”
pressed to Fig. 26.27(d). Actually, the documents in the com- regions to address this problem. Figure 26.27(a) does not have
pressed region have been further altered by using a semantic curvilinear distortions. Focuscontext visualizations can be
zooming function to give them a simplified visual form. The used as part of compact user controls. Keahey (2001) has cre-
form of Fig. 26.27(c) shows that this is essentially a step function ated an interactive scheme in which the bubble is used to “pre-
of two different slopes. An example of a two-dimensional step view” a region. When the user releases a button over the region,
[AQ51] function is the Table Lens (Fig. 26.28[a]). The Table Lens is a the system zooms in far enough to flatten out the bubble. Bed-
spreadsheet in which the columns of selected cells are ex- erson has developed a focuscontext pull-down menu (Beder-
panded to full size in X and the rows of selected cells are ex- son, 2000) that allows the viewing and selection of large lists of
panded to full size in Y. All other cells are compressed, and their typefaces in text editor Fig. 26.27(c).
content represented only by a graphic. As a consequence,
spreadsheets up to a couple orders of magnitude larger can be Perspective distortion. One interesting form of distort-
represented. ing visual transfer functions is 3D perspective. Although it can
By varying the visual transfer function (see, for example, the be described with a 2D distorting visual transfer function, it is
review by Leung and Apperley (1994), a wide variety of distorted usually not experienced as distorting by users due to the spe-
views can be generated. Figure 26.28(b) shows an application in cial perceptual mechanisms humans have for processing 3D.
which a visual transfer function is used to expand a bubble Figure 26.28(c) shows the Perspective Wall (Mackinlay, Robert-
around a local region on a map. The expanded space in the re- son, & Card, 1991). Touching any place on the walls animates
gion is used to show additional information about that region. its transition into the central focal area. The user perceives the
 26. Information Visualization • 539

FIGURE 26.28. Attention-Reactive Visualizations.

context area of the wall as an undistorted 2D image in a 3D tion to the focal area. Munzner (Munzner & Burchard, 1995) has
space, rather than as a distorted 2D image; however, the same extended this notion to 3D hyperbolic trees and used them to
sort of compression is still achieved in the nonfocus area. visualize portions of the Internet.

Alternate geometries. Instead of altering the size of

components, focuscontext effects can also be achieved by
changing the geometry of the spatial substrate itself. One ex- SENSEMAKING WITH VISUALIZATION
ample is the hyperbolic tree (Lamping & Rao, 1994). A visual-
ization such as a tree is laid out in hyperbolic space (which it- Knowledge Crystallization
self expands exponentially, just like the tree does), and then
projected on to the Euclidean plane. The result is that the tree The purpose of information visualization is to amplify cognitive
seems to expand around the focal nodes and to be compressed performance, not just to create interesting pictures. Informa-
elsewhere. Selecting another node in the tree animates that por- tion visualizations should do for the mind what automobiles do
 540 • CARD

for the feet. So here, we return to the higher level cognitive op- of acquiring information, capture, refers to acquiring informa-
erations of which information visualization is a means and a com- tion that is tacit or implicit. For example, when users browse the
ponent. A recurrent pattern of cognitive activity to which infor- World Wide Web, their paths contain information about their
mation visualization would be useful (though not the only one!) goals. This information can be captured in logs, analyzed, and vi-
is “knowledge crystallization.” In knowledge crystallization tasks, sualized (Chi & Card, 1999). It is worth making the point that ac-
there is a goal (sometimes ill-structured) that requires the ac- quiring information is not something that the user must neces-
quisition and making sense of a body of information, as well as sarily do explicitly. Search, monitoring, and capture can be
the creative formulation of a knowledge product, decision, or ac- implicitly triggered by the system.
tion. Examples would be writing a scientific paper, business or
military intelligence, weather forecasting, or buying a laptop
computer. For these tasks, there is usually a concrete outcome of Make sense of it. The heart of knowledge crystallization
the task—the submitted manuscript of a paper, a delivered brief- is sensemaking. This process is by no means as mysterious as it
ing, or a purchase. Knowledge crystallization does have charac- might appear. Because sensemaking involving large amounts
teristic processes, however, and it is by amplifying these that in- of information must be externalized, the costs of finding, orga-
formation visualization seeks to intervene and amplify the user’s nizing, and moving information around have a major impact on
cognitive powers. Understanding of this process is still tentative, its effectiveness. The actions of sensemaking itself can be ana-
but the basic parts can be outlined: lyzed. One process is extraction. Information must be got out
of its sources. In our hotel example, the hotel manager ex-
Acquire information. Make sense of it. Create something new. Act on it. tracted information from hotel records. A more subtle issue is
that information from different sources must be fused—that is,
registered in some common correspondence. If there are six
[AQ52] In Table 26.10, we have listed some of the more detailed ac-
called-in reports of traffic accidents, does this mean six different
tivities these entail. We can see examples of these in our initial
accidents, one accident called in six times, or two accidents re-
examples.
ported by multiple callers? If one report merely gives the county,
while another just gives the highway, it may not be easy to tell.
Acquire information. The FilmFinder is concentrated
Sensemaking involves finding some schema—that is, some de-
largely on acquiring information about films. Search is one of
scriptive language—in terms of which information can be com-
the methods of acquiring information in Table 26.10, and the
pactly expressed (Russell, Stefik, Pirolli, & Card, 1993). In our
FilmFinder is an instance of the use of information visualization
hotel example, permuting the matrices brought patterns to the
in search. In fact, Shneiderman (Card et al., 1999) has identi-
attention of the manager. These patterns formed a schema she
fied a heuristic for designing such systems:
used to organize and represent hotel stays compactly. In the
case of buying a laptop computer, the schema may be a table
Overview first, zoom and filter, then details-on-demand
of features by models. Having a common schema then permits
compact description. Instances are recoded into the schema.
The user starts with an overview of the films, and then uses
Residual information that does not fit the schema is noted and
sliders to filter the movies, causing the overview to zoom in on
can be used to adjust the schema.
the remaining films. Popping up a box gives details on the par-
ticular films. The user could use this system as part of a knowl-
edge crystallization process, but the other activities would take Create something new. Using the schema, information
place outside the system. The SmartMoney system also uses the can be reorganized to create something new. It must be orga-
TreeMap visualization for acquiring information, but this time nized into a form suitable for the output product and that prod-
the system is oriented toward monitoring, another of the meth- uct must be authored. In the case of the hotel example, the
ods in Table 26.10. A glance at the sort of chart in Fig. 26.5 allows manager created the presentation of Fig. 26.7(c).
an experienced user to notice interesting trends among the
hundreds of stocks and industries monitored. Another method
Act on it. Finally, there is some consequential output of
the knowledge crystallization task. That action may be to dis-
TABLE 26.10. Knowledge Crystallization Operators tribute a report or give a briefing, to act directly in some way,
such as setting up a new promotion program for the hotel or
Acquire Information Monitor
buying a laptop on the basis of the analysis, or by giving direc-
Search,
tives to an organization.
Capture (make implicit knowledge explicit)
Make sense of it Extract information
Fuse different sources
Find schema Levels for Applying Information Visualization
Recode information into schema
Create something new Organize for creation Information visualization can be applied to facilitate the vari-
Author ous subprocesses of knowledge crystallization just described.
Act on it Distribute It can also be applied at different architectural levels in a system.
Apply These have been depicted in Fig. 26.29. At one level is the use of [AQ53]
Act
visualization to help users access information outside the im-
 26. Information Visualization • 541

mediate environment—the infosphere—such as information on

the Internet or from corporate digital libraries. Figure 26.30(a) [AQ54]
shows such a visualization of the Internet (Bray, 1996). Web sites
are laid out in a space such that sites closer to each other in the
visualization tend to have more traffic. The size of the disk rep-
resents the number of pages in the site. The globe size repre-
sents the number of out-links. The globe height shows the num-
ber of in-links.
The second level is the information workspace. The infor-
mation workspace is like a desk or workbench. It is a staging
area for the integration of information from different sources.
An information workspace might contain several visualizations
related to one or several tasks. Part of the purpose of an infor-
mation workspace is to make the cost of access low for infor-
mation in active use. Figure 26.30(b) shows a 3D workspace for
the Internet, the Web Forager (Card, Robertson, & York, 1996).
Pages from the World Wide Web, accessed by users through
clicking on URLs or searches, appear in the space. These can
be organized into piles or books related to different topics. Fig-
ure 26.30 (c) shows another document workspace, STARLIGHT
(Risch et al., 1997). Documents are represented as galaxies of
points in space such that similar documents are near each other.
In the workspace, various tools allow linking the documents to
FIGURE 26.29. Levels of use for information visualization. maps and other information and analytical resources.

FIGURE 26.30. Information visualization applications.

 542 • CARD

The third level is visual knowledge tools. These are tools Information visualization is a set of technologies that use vi-
that allow schema forming and rerepresentation of information. sual computing to amplify human cognition with abstract infor-
The permutation matrices in Fig. 26.7, the SeeSoft system for mation. The future of this field will depend on the uses to which
analyzing software in Fig. 26.15(b), and the Table Lens in Fig. it is put and how much advantage it gives to these. Information
26.27(a) are examples of visual knowledge tools. The focus is on visualization promises to help us speed our understanding and
determining and extracting the relationships. action in a world of increasing information volumes. It is a core
The final level is visually enhanced objects, coherent infor- part of a new technology of human interfaces to networks of de-
mation objects enhanced by the addition of information visual- vices, data, and documents.
ization techniques. An example is Fig. 26.30(d), in which voxel
data of the brain have been enhanced through automatic sur-
face rendition, coloring, slicing, and labeling. Abstract data
structures representing neural projects and anatomical labels ACKNOWLEDGMENT
have been integrated into a display of the data. Visually en-
hanced objects focus on revealing more information from some This chapter benefited from detailed and generous discussion with
object of intrinsic visual form. Jock Mackinlay and tutorial collaborations with Ed Chi of PARC.

[AQ58] References
Ahlberg, C., & Shneiderman, B. (1994a). Visual information seeking using Chi, E. H., & Riedl, J. T. (1998). An operator interaction framework for
the filmfinder. Paper presented at the Conference Companion of visualization spreadsheets. Paper presented at the Proceedings of
CHI’94, ACM Conference on Human Factors in Computing Systems. InfoVis’98, IEEE Symposium on Information Visualization.
Ahlberg, C., & Shneiderman, B. (1994b). Visual information seeking: Chuah, M. C., Roth, S. F., Mattis, J., & Kolojejchick, J. A. (1995). Sdm:
Tight coupling of dynamic query filters with starfield displays. Malleable information graphics. Paper presented at the Proceed-
Paper presented at the Proceedings of CHI’94, ACM Conference on ings of InfoVis’95, IEEE Symposium on Information Visualization,
Human Factors in Computing Systems, New York. New York.
Battista, G. D., Eades, P., Tamassia, R., & Tollis, I. G. (1994). Annotated Cleveland, W. S., & McGill, M. E. (1988). Dynamic graphics for statistics.
bibliography on graph drawing. Computational Geometry: Theory Pacific Grove, California: Wadsworth and Brooks/Cole.
and Applications, 4(5), 235–282. Cruz, I. F., & Tamassia, R. (1998). Graph drawing tutorial. Retrieved
Becker, R. A., Eick, S. G., & Wilks, A. R. (1995, March 1). Visualizing net- (date, from http://www.cs.brown.edu/people/rt/papers/gd-tutorial/
work data. IEEE Transactions on Visualization and Computer gd-constraints.pdf.
Graphics, 1, 16–28. Eick, S. G., Steffen, J. L., & Sumner, E. E. (1992, November). Seesoft—a
Bederson, B. B. (2000). Fisheye menus. Paper presented at the UIST tool for visualizing software. IEEE Transactions on Software Engi-
2000, ACM Symposium on User Interface Software and Technology neering, 18(11), 957–968.
(CHI Letters, 2(2)). Eick, S. G., & Wills, G. J. (1993, October 25–29). Navigating large net-
Bertin, J. (1983). Semiology of graphics: Diagrams, networks, maps works with hierarchies. Paper presented at the Proceedings of IEEE
(W. J. Berg, Trans.). Madison, WI: University of Wisconsin Press. Visualization’93 Conference, San Jose, CA.
(Original work published 1967) Engelhardt, Y., Bruin, J. D., Janssen, T., & Scha, R. (1996). The visual
Bertin, J. (1981). Graphic constructions (P. Scott, Trans.). In Graphics grammar of information graphics. In S. University (Ed.), Artificial
constructions and graphic information-processing (pp. 24–31). intelligence in design workshop notes (pp. 24–27).
Berlin, Germany: Walter De Gruyter. (Original work published 1977) Fairchild, K. M., Poltrock, S. E., & Furnas, G. W. (1988). Semnet: Three-
Bray, T. (1996). Measuring the web. Computer Networks and ISDN Sys- dimensional representations of large knowledge bases. In R. Guin-
tems, 28(7–11-May), 992. don (Ed.), Cognitive science and its applications for human-
Card, S. K., Mackinlay, J. D., & Shneiderman, B. (1999). Information computer interaction (pp. 201–233). Hillsdale, New Jersey: Lawrence
visualization: Using vision to think. San Francisco: Morgan Kauf- Erlbaum Associates.
mann Publishers. Feiner, S., & Beshers, C. (1990). Worlds within worlds: Metaphors for
Card, S. K., Moran, T. P., & Newell, A. (1986). The model human proces- exploring n-dimensional virtual worlds. In ACM symposium on User
sor: An engineering model of human performance. In J. Thomas Interface Software.
(Ed.), Handbook of perception and human performance (pp. 41– Freeman, E., & Fertig, S. (1995). Lifestreams: Organizing your electronic
35). New York: John Wiley and Sons. life. Paper presented at the Proceedings of AAAI Fall Symposium on
Card, S. K., Robertson, G. G., & York, W. (1996). The webbook and the AI Applications in Knowledge Navigation.
web forager: An information workspace for the world-wide web. Furnas, G. W. (1981). The fisheye view: A new look at structured files.
Paper presented at the Proceedings of CHI’96, ACM Conference on In B. Shneiderman (Ed.), Readings in information visualization:
Human Factors in Computing Systems, New York. Using vision to think (pp. 312–330). San Francisco: Morgan Kauf-
Casner, S. (1991, April 2). Task-analytic approach to the automated mann Publishers, Inc.
design of graphic presentations. ACM Transactions on Graphics, 10, Healey, C. G., Booth, K. S., & Enns, J. T. (1995). High-speed visual esti-
111–151. mation using preattentive processing. ACM Transactions on Com-
Chi, E. H., & Card, S. K. (1999). Sensemaking of evolving web sites puter-Human Interaction, 3(2), 107–135.
using visualization spreadsheets. Paper presented at the Infovis Inselberg, A. (1997). Multidimensional detective. Paper presented at the
1999, IEEE Conference on Information Visualization 1999, San Proceedings of InfoVis’97, IEEE Symposium on Information Visu-
Francisco. alization, IEEE Information Visualization.
 26. Information Visualization • 543

Inselberg, A., & Dimsdale, B. (1990). Parallel coordinates : A tool for Robertson, G. G., Card, S. K., & Mackinlay, J. D. (1989). The cognitive co-
visualizing multi-dimensional geometry. Paper presented at the Pro- processor for interactive user interfaces. Paper presented at the
ceedings of IEEE Visualization’90 Conference, Los Alamitos, CA. Proceedings of UIST’89, ACM Symposium on User Interface Soft-
Keahey, T. A. (2001, October 22–23, 2001). Getting along: Composition ware and Technology.
of visualization paradigms. Paper presented at the Infovis 2001, Robertson, G. G., Card, S. K., & Mackinlay, J. D. (1993). Information visu-
IEEE Information Visualization 2001, San Diego, California. alization using 3d interactive animation. Communications of the
Keim, D. A., & Kriegel, H.-P. (1994, September). Visdb: Database explo- ACM, 36(4), 57–71.
ration using multidimensional visualization. IEEE Computer Graph- Robin, H. (1992). The scientific image: From cave to computer. New
ics and Applications, 40–49. York: H. N. Abrams, Inc.
Kosslyn, S. M. (1994). Image and brain: The resolution of the imagery Roth, S. F., & Mattis, J. (1990). Data characterization for intelligent
debate. Cambridge, MA: The MIT PRess. graphics presentation. Paper presented at the Proceedings of
Lamping, J., & Rao, R. (1994). Laying out and visualizing large trees CHI’90, ACM Conference on Human Factors in Computing Systems,
using a hyperbolic space. Paper presented at the Proceedings of New York.
UIST’94, ACM Symposium on User Interface Software and Technology. Russell, D. M., Stefik, M. J., Pirolli, P., & Card, S. K. (1993). The cost struc-
Leung, Y. K., & Apperley, M. D. (1994, June). A review and taxonomy of ture of sensemaking. Paper presented at the Proceedings of INTER-
distortion-orientation presentation techniques. ACM Transactions CHI’93, ACM Conference on Human Factors in Computing Systems,
on Computer-Human Interaction, 1(2), 126–160. Amsterdam.
Lin, X., Soergel, D., & Marchionini, G. (1991). A self-organizing seman- Shneiderman, B. (1992). Tree visualization with tree-maps: A 2-dimensional
tic map for information retrieval. Paper presented at the Proceed- space filling approach. ACM Transactions on Graphics, 11(1), 92–99.
ings of SIGIR’91, ACM Conference on Research and Development Shneiderman, B., & Wattenberg, M. (2001). Ordered tree layouts. Paper
in Information Retrieval, Chicago, IL. presented at the IEEE Symposium on Information Visualization, San
MacEachren, A. M. (1995). How maps work. New York: The Guilford Diego, California.
Press. Spence, R. (2000). Information visualization. Harlow, England: Addi-
Mackinlay, J. D. (1986a). Automatic design of graphical presentations. son-Wesley.
Unpublished doctoral dissertation, Stanford University, California. Spence, R., & Apperley, M. (1982). Data base navigation: An office envi-
Mackinlay, J. D. (1986b). Automating the design of graphical presenta- ronment for the professional. Behavior and Information Technol-
tions of relational information. ACM Transactions on Graphics, 5(2), ogy, 1(1), 43–54.
110–141. Tamassia, R. (1996, December 4). Strategic directions in computational
Mackinlay, J. D., Robertson, G. G., & Card, S. K. (1991). The perspective geometry working group report. ACM Computing Surveys, 28,
wall: Detail and context smoothly integrated. Paper presented at 591– 606.
the Proceedings of CHI’91, ACM Conference on Human Factors in Tufte, E. R. (1983). The visual display of quantitative information.
Computing Systems, New York. Cheshire, CT: Graphics Press.
McCormick, B. H., & DeFanti, T. A. (1987, November 6). Visualization is Tufte, E. R. (1990). Envisioning information. Cheshire, CT: Graphics
scientific computing. Computer Graphics, 21. Press.
Munzner, T., & Burchard, P. (1995, December 14–15). Visualizing the Tufte, E. R. (1997). Visual explanations: Images and quantities, evi-
structure of the world wide web in 3d hyperbolic space. Paper pre- dence and narrative. Cheshire, CT: Graphics Press.
sented at the Proceedings of VRML ’95. Tukey, J. W. (1977). Exploratory data analysis. Reading, MA: Addison-
Playfair, W. (1786). The commercial and political atlas. London. Wesley.
Resnikoff, H. L. (1989). The illusion of reality. New York: Springer- Tweedie, L. A., Spence, R., Dawkes, H., & Su, H. (1996). Externalising
Verlag. abstract mathematical models. Paper presented at the Proceed-
Risch, J. S., Rex, D. B., Dowson, S. T., Walters, T. B., May, R. A., & Moon, ings of CHI’96, ACM Conference on Human Factors in Computing
B. D. (1997). The starlight information visualization system. Paper Systems.
presented at the Proceedings of IEEE International Conference on Ware, C. (2000). Information visualization: Perception for design. San
Information Visualization, London, England. Francisco: Morgan Kaufmann Publishers.
Author Notes: [AQ1]Figure cite: Fig. 27.1.
AU Note 1: Ahlberg, C., & Shneiderman, B. (1994a) please provide date [AQ2]Figure cite: Fig. 27.2.
of conference (month and day) and location of conference. [AQ3]Figure cite: Fig. 27.3.
AU Note 2: Ahlberg, C., & Shneiderman, B. (1994b)) please provide date [AQ4]Figure cite: Fig. 27.4.
of conference (month and day).
[AQ5]FN1, p. 54.
AU Note 3: Bederson, B. B. (2000). Is this a published document? Please
clarify. If published please use the APA style guide format for regu-
[AQ6]Figure cite: Fig. 27.5.
larly published proceedings for this reference on p. 259, section [AQ7]”scroll”?
4.16. [AQ8]Figure cite: Fig. 27.6.
AU Note 4: Card, S. K., Robertson, G. G., & York, W. (1996). Please pro- [AQ9]”scrolls”?
vide date of conference. [AQ10]”analysts’”?
AU Note 5: Chi, E. H., & Card, S. K. (1999). please provide date of con- [AQ11]Table cite: Table 27.1.
ference (month and day). [AQ12]AU: Are italics necessary here?
AU Note 6: Chi, E. H., & Riedl, J. T. (1998). Please provide date of con- [AQ13]”into more compact forms”? “into a more compact
ference (month and day) as well as location. form”?
AU Note 7: Chuah, M. C., Roth, S. F., Mattis, J., & Kolojejchick, J. A.
[AQ14]”physically-rather than visually-based.”
(1995). Please provide date of conference.
AU Note 8: Cruz, I. F., & Tamassia, R. (1998). Please provide date
[AQ15]Table cite: Table 27.2.
retrieved. [AQ16]Figure cite: Fig. 27.8.
AU Note 9: Engelhardt, Y., Bruin, J. d., Janssen, T., & Scha, R. (1996). [AQ17]Figure cite: Fig. 27.9.
Please provide publisher information for this reference. [AQ18]”their combinations?”
AU Note 10: Feiner, S., & Beshers, C. (1990). Is this a published docu- [AQ19]Figure cite: Fig. 27.10.
ment? If so please provide publisher information. [AQ20]”through”?
AU Note 11: Freeman, E., & Fertig, S. (1995). Please provide dates of [AQ21]Table cite: Table 27.3.
conference as well as location. [AQ22]”June”?
AU Note 12: Inselberg, A. (1997). Please provide dates of conference as [AQ23]”June”?
well as location.
[AQ24]Table cite: Table 27.4.
AU Note 13: Inselberg, A., & Dimsdale, B. (1990). Please provide dates of
conference.
[AQ25]Table cite: Table 27.5.
AU Note 14: Lamping, J., & Rao, R. (1994). Please provide dates of con- [AQ26]Figure cite: Fig. 27.11.
ference as well as location. [AQ27]Figure cite: Fig. 27.12.
AU Note 15: Lin, X., Soergel, D., & Marchionini, G. (1991). Please pro- [AQ28]Figure cite: Fig. 27.13.
vide dates of conference. [AQ29]Table cite: Table 27.7.
AU Note 16: Mackinlay, J. D., Robertson, G. G., & Card, S. K. (1991). [AQ30]Table 27.7 comes before Table 27.6.
Please provide location of conference. [AQ31]Table cite: Table 27.6.
AU Note 17: McCormick, B. H., & DeFanti, T. A. Please provide page [AQ32]Table cite: Table 27.8.
numbers. [AQ33]Figure cite: Fig. 27.14.
AU Note 18: Munzner, T., & Burchard, P. (1995, December 14-15). Please
[AQ34]Figure cite: Fig. 27.15.
provide location of conference.
AU Note 19: Playfair, W. (1786). Please provide publisher.
[AQ35]Figure cite: Fig. 27.16.
AU Note 20: Risch, J. S., Rex, D. B., Dowson, S. T., Walters, T. B., May, R. [AQ36]Figure cite: Fig. 27.17.
A., & Moon, B. D. (1997). Please provide dates of conference. [AQ37]FN2, p. 54.
AU Note 21: Robertson, G. G., Card, S. K., & Mackinlay, J. D. (1989). [AQ38]”searches”? “traverses”?
Please provide location and dates of conference. [AQ39]Figure cite: Fig. 27.18.
AU Note 22: Roth, S. F., & Mattis, J. (1990). Please provide dates of con- [AQ40]Figure cite: Fig. 27.20.
ference. [AQ41]FN3, p. 54.
AU Note 23: Russell, D. M., Stefik, M. J., Pirolli, P., & Card, S. K. (1993). [AQ42]”creates”?
Please provide dates of conference. [AQ43]Figure cite: Fig. 27.21.
AU Note 24: Shneiderman, B., & Wattenberg, M. (2001). Please provide
[AQ44]Figure cite: Fig. 27.22.
dates of conference.
AU Note 25: Tweedie, L. A., Spence, R., Dawkes, H., & Su, H. (1996).
[AQ45]Figure cite: Fig. 27.23.
Please provide dates and location of conference. [AQ46]Table cite: Table 27.9.
[AQ47]Figure cite: Fig. 27.24.
[AQ48]Figure cite: Fig. 27.25.
[AQ49]Figure cite: Fig. 27.26.
[AQ50]Figure cite: Fig. 27.27.
[AQ51]Figure cite: Fig. 27.28
[AQ52]Table cite: Table 27.11.
[AQ53]Figure cite: Fig. 27.29.
[AQ54]Figure cite: Fig. 27.30.
[AQ55]FN1, p. 6.
[AQ56]FN2, p. 31.
[AQ57]FN3, p. 32.
[AQ58]AU: Please see color coded notes at the bottom of page.

View publication stats