You are on page 1of 7

R

1---
Human
Factors
inComputiig
Systems WI ’940 “Celebrating
hwdepemfenw”

The Cost-of-Knowledge Characteristic Function:


Display Evaluation for Direct-Walk Dynamic
Information Visualizations
Stuart K. Card, Peter Pirolli, and Jock D. Mackinlay

Xerox Palo Alto Research Center


3333 Coyote Hill Road
Palo Alto, California 94304
E-mail: card@parcxerox.tom,
piroUi@parc.xerox.tom,
mackinlay@parc.xerox. com

ABSTRACT Information retrieval and other information-handling


In this paper we present a method, the Cost-of-Knowledge systems reorganize this cost structure of information
Characteristic Function, for characterizing information relative to some task. For example, retrieving paper
access from dynamic The paper works out this
displays. documents from filing cabinets and placing them on a desk
method for a simple, but important, class of dynamic reduces the time costs substantially for a task in which the
displays called direct-walk interactive information documents must be repeatedly referenced.
visuatizations, in which information is accessed through a
sequence of mouse selections and key selections. The A goal in designing information access systems is to
method is used to characterize a simple calendar task for an rearrange this cost structure in beneficial ways. In previous
application of the Information Visualizer, to compute the papers, we have reported the designs of experimental
changes in characterization as the result of possible program programming systems whose inkx-faces were designed for
variants, and to conduct empirical comparison between this purpose [3] [4]. But methods are needed for
different systems with the same function. conceptualizing and measuring the abilities of these and
other systems to bring about the desired result.
KEYWORDS: Information visualization, dynamic
displays, methodology, evaluation, 3D user interfaces, In this paper we propose an abstraction, the Cost-of-
Information Visualizer. Knowledge Characteristic Function, for characterizing the
effect of the design of a dynamic display or human-
INTRODUCTION computer dialogue on the cost structure of information.
The personal computer is changing from a device into The purpose of the paper is to work out and measure
which information is mainly put for authoring or analysis empirically this abstraction for a simple case of information
(e.g., desktop publishing) to a device from which access, “direct-walk” information visualizations, by which
information is mainly accessed (e.g., CD-ROM we mean the use of the mouse to point to and gesture over
encyclopedias and on-line data services). This trend can displays of an information structure so as to navigate from
only increase as the national data superhighway comes into one place in that structure to another.
being and as local memory costs continue to drop. Tennant
and Heilmeier [8] estimate, for example, that the amount of THE COST-OF-KNOWLEDGE CHARACTERISTIC
information available through one’s computer by 1995 will FUNCTION
be more than 10,000 times greater than the information We have argued that, at least in a world of abundant
available at the time of their writing, about 1991. A information, but scarce time, the fundamental information
challenge of the 1990s is to develop ways of using access task is not finding information, but the optimal use
emerging technologies to manage the complexity inherent of a person’s scarce time in gaining information [7]. That
in accessing and utilizing such vast quantities of is, the important thing is to maximize information benefits
information. per unit cost (The unit of cost considered in this article is
primarily the user’s time). To aid in doing this, we need to
A key observation is that information in an information know how much additional information becomes available
system has a cost structure, that is, a set of different costs for each additional amount of time expended. We call the
for the information in different parts of the system [3]. curve this notion defines the Cost-of-Knowledge
Characteristic Function.
Permission to copy without fee all or part of this material is
granted provided that the copies are not made or distributed for
Fig. 1 shows a schematic plot of this function. Curve A
direct commercial advantage, the ACM copyright notice and the
shows a hypothetical office in which information is
title of the publication and its date appear, and notice is given
that copying is by permission of the Association for Computing
hierarchically arranged: Small amounts of information are
Machinery. To copy otherwise, or to republish, requires a fee placed on the low-cost access desktop, larger amounts of
and/or specific permission.
CH194-4/94 Boston, Massachusetts USA
e 1994 ACM 0-89791 -650-6 {9410238 . ..$3.50

238
Boston,
Massachusetts
USA* April24-28,1994 Human
Factors
in Compu[ing
Sys[ems
II%?
information are placed in the more expensive-to-access, but develop the methodology. In the remainder of this paper,
more capacious, desk file drawer, and large quantities of we perform a series of analyses as exercises to establish the
information are in the file cabinet. For simplicity, it is feasibility and utility of this concept. It should be noted
assumed that the average time to access the information that the emphasis is on the development of the
within each of these categories is the same, and we have methodology itself and that while useful information may
ignored the staircase function produced by the fact that the be revealed about the systems we analyze (after all, that is
repositories are of discrete sizes. the purpose!), our analyses do not constitute a complete
evaluation of these systems, since that would require
CURVE B considering what other figures of merit might be relevant as
~c~ well.
IMPROVEMENT)
EMPIRICAL CHARACTERIZATION OF A SYSTEM
In our first analysis, we empirically measure the function of
Fig. 1 for one particular task done by users of the Spiral
Calendar, an application of the Information Visualizer [6]
(HIERARCHICAL developed by Mackinlay and DeLine [5]. A user can access
the schedule for various calendar dates by selecting objects
that represent the appropriate period, then selecting the unit
within that period, and so on. For example, to select June
4, 1982, the user would select the year 1982 within the
Decade 1980-1989 object, this would cause the Year 1982
TIME COST (S)
object to fly up and grow large. Then the user would select
Fig. 1. Cost of Knowledge Characteristic Function June from that Year object. This would cause the June
Month object to fly up. The user would select 4 from the
Now suppose that we were able to invent some device or June month object causing the containing Week object to
procedure that improves information access. The Cost of fly up. Finally, the user would select 4 from the week
Knowledge Characteristic Function should show that object, causing the daily schedule for that day to appear.
improvement by having at least some portion above the There could be fewer selections or more selections
original curve (e.g., Curve B in Fig. 1). Notice that we can depending on the condition of the display from which the
harvest this benefit in two ways, as shown by the arrows, user started.
If we keep the time cost the same, we could access more
documents (arrow a). On the other hand, if we keep the
number of documents the same, we can access them for a
lower cost (arrow b). In this way, the Cost-of-Knowledge
Characteristic Function is intended to help us reason
through more complicated consequences of system
improvements than just thinking that one system is better
than another.

Direct Manipulation Walk of Information


Structures
We now attempt to measure the function in Fig. 1 for an
actual system. To keep the analysis simple, we investigate
a basic, but important, information access task, which we
dub a direct walk of an information structure.

We define a direct walk to be a task in which a user


navigates from a starting point to a goal point in an
information structure by a series of mouse points or
other direct-manipulation methods. Examples would Fig. 2. The Spiral Calendar.
be the series of mouse clicks and button choices
This system was measured using the following procedure:
required to operate the Macintosh hierarchical file
system or a typical HyperCard stack or many help Task
systems. The essence of a direct-walk is that an The general task measured can be described as follows:
information structure is displayed and the user points
to, flies to, or gestures over some part of this visible A user is looking at the &ta’led hourly calendar for a
structure resulting in a new display at which time the certain day, and he or she wishes to view the daily
cycle is repeated until the goal is found. calendar for another day. How long will it take to do
this?
We do not believe the Cost of Knowledge Characteristic
Function is limited to this class of dialogues. We simply
Users had to position the calendar to a set of 11 different
choose these dialogues as a simple case on which to days related to the starting date (chosen to be September 7,

239
ml HumanFactorsinComputingSystems
,,, ,,.
CHI’94 * “Celebra/i~]:
I/l(erdepetrde/ire”

1993). These days were chosen to lie on a logarithmic The set of 11 tasks was randomized into a block of trials.
scale We specifically chose such a large range because we Each user performed 5 of these blocks, each block
want to understand how the interaction scales with size. separately randomized, for a total of 55 trials/user. Users
were allowed to take a break between blocks if they wished.
Users On each trial, the user flipped a page in a notebook asking
The measurement was done on four users, members of the him or her to navigate through the calendar to the day
professional staff of the lab. Users varied in their display of a specific date. The trials were videotaped and the
experience. Two had never used the Information Visualizer, time measured from when the user had turned from the
one was one of the designers of the Spiral Calendar. notebook and was facing the display until the day page was
done displaying.
Procedures
As a warm-up, the user first performed a set of 11 accesses Empirical Results
to dates different from, but similar to in time, those The results of the measurement are in Table 1, column (4).
actually used in the experiment. These served to help the Plotting time to access a date as a function of number of
user assimilate the procedures of the experiment, to learn days back (column 2) gives Fig. 3. As might be expected,
how to operate the Spiral Calendar, and to ask questions. the time required increases with the number of days back.
As a limitation on the prototype, only data from 1993 was This is obviously because there are more steps in the
actuaIIy contained in the database and the calendars said dialogue to reach distant dates. How much additional time
1993, regardless of the simulated year. Users were simply can be understood by analyzing the predominant method
told of this limitation and that this data was being used to users utilized.
simulate a larger range of years.

Table 1. Computation of Cost of Knowledge Characteristic Function for Spiral Calendar.

(1) (2) (3) (4) 5) (6)


TASK DAYS DATE ACCESS TIMEa ME$HOD No.
BACK MEAN HD
(s) c!%
~
1 1 Sep. 6, 1993 5 .Ln n. n... 1

-+=T%m‘h ;T
2 3 Sep. 4, 1993 11 T
2 In A,, (T ‘7$2 100’2 1A

. .,”, ’...~. “,. ,,.,

5 100 May 30, 1993 , . . ..–-. .- ,, , ?


6 300 NOV. 11, 1992 T~ fan 20 II v... I .A I 179+n7<i 17< I 19 I ‘IA<
.“. ”L”. J.
J
7 1000 Dec. 12, 1990 17.8*0.35
8 3000 Seu. 25. 1982 21.1+0.28 T
9 10,000 ipr~ i; 1966 21 .2+z0.69
10 30,000 . . . 1.
Jul. 20, IQl 1
9n
-- 7+0 ..-A6
.,--
11
t 1 1 I I
* ,

11 100,000 NOV. 23, 1719


_._. , 7.4 . . .40
. ..‘3+1 . . -------
I Cenhrv i (i I ‘24.3+ 1.5 24.6 10 I _-, ——
36.525-
300,000 II Millennium 7 II I 28.1 I 1 0 I 365,250
I 1,000,000 I I II Era I 8 I I 31.6 ] 10 I 3,652,500
a Each mean is based on 4 users x 5 repetitions = 20 data points
b Computed using Time= 3.346 + 3.535 * NCycles

Methods = Decadedisplsy
GET-YEAR ..0 if neceess
The cost in time for accessing some date can be GOAtJ SELECT-YEAR: (17197
characterized in terms of the major methods available to POINT-TO (171 9))
users. Let us take the most extreme case, accessing the date *Year-display
G ET. MONTH . . . if nmesaary
November 23, 1719. A GOMS model [2] for this GOAL: SELECT-MONTH: (November)
procedure would be
‘O’NT-TO(No~fiRtl-clisplay
GET-DAY . . . if n*ssar-y
CENTURY-METHOD =
GOAL: SELECT-WEEK: [??]
GOAL: DO-TASK POINT-TO [23]
GOAL: GET-DATE a Week-display
TURN-TOrMANUSCRIPTl
GET-DAY . . . if necessary
GET.DATE GOAL: SELECT-DAY: [23]
GOA&E+~~~~S-DAY-CALEN DAR POINT-TO (23))
. . . . if n~sqry
+ DaydispJay
GOAL: SELECT-CENTURY (1700’s)
POINT-TO (Century=1700-1790s))
= Century-display Neglecting the initial part of this method that has to do
GET-YEAR . . . if necessary
with our experimental procedure, the method can be
GOAL: SELECT-DECADE (171 (t’s)
POINT-TO (1710-1719)) summarized in terms of seven cycles of pointing and

240
Boston,
Massachusetts
USA* April24-28,1994 Human
Factors
in Compu[ig
Sys[ems
R
L—–
display+me each for Century, Deeade, Year, Month, Week, accessible (column 10) as a function of the cost in time
and Day. Other methods used in our measurement are the (column 8) in Table 1. This is done in Fig. 5 (Curve A:
same, except that the larger units of time, such as the Spiral Calendar). This graph, the first actual calculation of
century, or the decade, or even the month, are eliminated if this concept we have been able to achieve, shows how, as
the date is close enough, To make the methods easy to talk would be expected in a reasonable system, the small
about, we name them by the largest unit of time selected. amounts of knowledge can be accessed quickly, larger
The method for each task is listed in column (5) of Table 1. amounts of knowledge rquire longer times. For simplicity
we have omitted the stair-case detail of the curve that would
track abrupt representation shifts (e.g., from month to
year). The metric is roughly linear, in semi-log
30
coordinates, indicating that the items accessed increase
T Century exponentially. This linear shape of the curve may be a
natural form for describing accessibility y with cost for well-
designed systems.

o~
1 100 10000 1000000
a o~
o 2 4 6

NUMBER OF SELECT-DISPLAY
DAYS BACK CYCLES
Fig. 3. Time as a function of number of days back.

Model for Access Time Fig. 4. Time as a function of the number of selection-
On Fig. 3, we have circled those data points that are done display cycles.
with the same method and collapsed the data cells to one
data point per method in Table 1, columns (5) and after. At COMPUTATIONAL CHARACTERIZATION OF
this point, we can use our data to fit a simple model in DESIGN IMPROVEMENTS
order to characterize the direct walk time. To a first order of With the Cost-of-Knowledge Characteristic Function
analysis, the time (in seconds) to select a date is just conceptually in hand, we can now use it to help reason
proportional to the number of cycles required (l?ig. 4). By a about and discover variants in the system design. Let us
regression analysis, discuss briefly the effect of some design changes on the
system measured in the last section.
Time to Access= 3.3 + 3.5 * NCycles. (1)
One possibility is to speed up the system response. We
This model allows us to give a smoother characterization notice (from Eq. 1) that the time per picking cycle is on the
than the individual data points and we list the model times order of 3.5 s. This would seem to be relatively long.
in Table 1, column (8). Assuming a mouse point of around 1 s [2] and a response
animation time of 1 s [3] suggests that 2 s/cycle should be
Number of elements accessed possible The discrepancy suggests re-examination of the
Finally, we compute the number of days that it is possible animation algorithms< We now replot (as recalculated from
to reach in less than or equal to a certain amount of time. Eq. 1) the Cost of Knowledge Characteristic Function that
This is done by computing the number of days existing
would have resulted from a faster 2 second user-action cycle
within the different periods serviced by the methods in (Curve Bin Fig. 5). The curve is tilted upwards indicating
Table 1. We list the result in Table 1, column (10). With
an improvement in the system (if it can be achieved
the model, we can make reasonable estimates of what the
computationally and if some other phenomenon does not
data would be for other dates not actually measured (e.g., intervene). Notice that in this case we have plugged the
1,000,000 days distant). results from previous models into our new model to cascade
Cost-of-Knowledge Characteristic Function the speed with which we can think about design variants.
We can now plot the Cost-of-Knowledge Characteristic
Another possibility is to eliminate the Week display (or
Function of Fig. 1 by plotting the number of elements
probably better, to integrate it into the same display as the

241
Human
Factors
in Computig
Systems

10000000
T C: No Week

F-
(n 1000000
0
0

100000

w 10000
A
m
z
w
u)
1000
Ill
0
0
a 100

10
1=
I
~p u ,
1 I I
1 i
I
1
o 20 40 60 80 100 120

COST (S)

Fig. 5. Cost of Knowledge Characteristic Function for several variants of the Spiral Calendar and the CM calendar program.

Day) in order to reduce the number of action-display cycles Knowledge Characteristic Function Fig. 5 grows a bump at
required. If we were to do just this, then the Cost of the bottom (Curve C), because of the larger branching
factor at the Month level. In Curve D, we combine both
Table 2. Computation of Cost of Knowledge Characteristic variants.
Function for Sun CM Calendar.
With the Cost of Knowledge Characteristic Function, we
are led to view variants in terms of their effect on the cost-
structure of access rather than just on a single point.

EMPIRICAL COMPARISON AMONG SYSTEMS


In addition to using calculations to do paper comparisons
I I I 1 , among system variants before they are built, we cart also
1 2.4~0.56 Month 2.3*0.34 2.8 30 use the Cost of Knowledge Characteristic Function to do
comparisons between competing systems. AS an example,
---i2 2.2&0.25 II II Ii II we repeat our measurement and calculation, but this time
I for the Sun calendar program CM.
3 5.0+0.60 I A40nth2 I 4.6+0.14 I 4.2 I 61
Procedure
4 4.3*().53 The tasks and the procedure were exactly the same as for the
Iris Spiral Calendar above. Four users participated in this
5 9.0*0.90 Year 9.0*0.55 8.1 365
measurement, two of these were users measured in the
I I I I 1
8.7+0.90 4.5 731 previous Spiral Calendar. The results are shown in Table
6 8.7+0.90 I Year2
I I I 2, comparable to Table 1.
I I
7 10.03+0.80 I Year4 I 10.0+0.80 I 10.2 I 1461
The GOMS analysis [2] of the methods is summarized in
I I I I I
short-hand form in Table 3 column (4). In this analysis,
8 I 12.8~0.70 Yearn 12.8*0.70 12.8 4018
I I I I I there are three operatorx
a Each mean is based on 4 users x 5 repetitions = 20 data
m point, menu pull-down, and select
points.
P point and select
b Time = 1.340 + 3.889 m + 1.412 P + 0-.362 B.
B press a button (not including pointing)

242
Boston,
Massachusetts
USA* April24-28,1994 Human
Factors
in Computing
Systems
!!!Ei!
These operators include the system response time of this PROBABILITY DENSITY FUNCTION
particular system. They do not count the mental Fimlly, it should be noted that whereas we have determined
preparation time [2]. The time taken by the different the Cost of Knowledge Function, we have not as yet
methods is given in Table 3, which analyzes the weighted the items accessed by their value. There are
predominant method used. (Remember, methods were several such weighings, but one of the most important is
restricted to be direct walk). to weight items by a probability density function describing
their frequency of use.
Table 3. Analysis of methods for CM calendar task.
Anderson and Schooler [1] have shown that for many
ACIION ANALYSIS different kinds of information (e.g., news articles in the
Select Date P New York Times or messages in electronic mail) the
display probability an item D days old will be needed is given by
Month2 Month Select PR.EV button 2P
Pr(neededlll days ago) = A /(A+ De), where
_ ChJ)hJ_
Select Day
where A and C are constants. For the case of electronic
Year Month Select year on m+2P mail, A = .34 and C = 0.83, hence
display VIEW pulldown
menu Pr{neededlD days ago) = 0.34/ (0.34+ D0.83).
Year Select Month
Month Select Day If we multiply this function with the Cost of Knowledge
Year(n) Month Select Year on m+ 3P+ Characteristic Functions in Fig. 5 to obtain a curve
display VIEWpulldown (n-2)B expressing the expscted cost of accessing different numbers
menu of items (see Fig. 6). In the case of the calendars, it
Year Select PREV button expresses the fact that the user is likely to access recent
(n-1) times dates much more frequently. The area under the curves is
Year Select Month related to the total costs of using the two programs and the
Month Select Day curve also shows in what area the costs are concentrated.
Fig. 6 shows that the expected cost for the user is heavily
In this case, we are not trying to predict method times, but contained in the in the most recent hundred days. For this
to analyze them, so, as before we use regression analysis to reason, the Spiral Calendar prototype tested would be more
assign numbers to the operators. The regression gives expensive to use than CM (if only the direct walk feature
were considered). On the other hand, if the task involved
Time = 1.3 + 3.9 m + 1.4 P + 0.36 B . (2) reference to historical dates, then a different probability
density function would be appropriate.
This equation (the equivalent of Eq. 1 for CM) is used to
determine a smoother version of the Cost of Knowledge
Characteristic Function in Table 2, column (5). Finally, 1.8~
the number of days accessible within a given iso-cost 1,6
contour is determined from an inspection of the program 1.4
displays and summarized in Table 2, column (6).
before, we plot number of items accessible (column 6)
against cost (column 5) in Curve E in Fig. 5.
As

A comparison between the Cost of Knowledge functions in


3jl.2 .i
l-l
~
0
0.8

0.6
0.4
\
A ‘
\
Spiral Calendar

Fig. 5 shows a rather dramatic contrast. The Spiral


Calendar uses a uniform direct-walk method to access all
dates and costs go up logarithmically. The CM program is 1 100 10000 1000000 100000000
at an advantage for lower numbers of items, but costs
radically increase for numbers of items over about 10,000.
(Remember, however, we are not characterizing the NUMBER OF DAYS ACCESSED
programs themselves, only certain methods. For example,
we are not considering methods such as typing in the date
directly that would be useful for larger numbers of items.) Fig. 6. Expected probability-weighted costs
associated with retrievals.
The Spiral Calendar suffers because of its high cost
intercept. Clearly there is payoff in concentrating effort at CONCLUSION
the low end of the curve to shift it to the left. The CM This paper has introduced the Cost-of-Knowledge
Characteristic Function as a method for analyzing
program suffers because users must shift through a more
complex space of methods, changing methods for different interactive information visualizations. The metric was
measured for direct-walk dialogue taken from two calendar
regions of the space.
programs. We have also done calculations showing the

243
m , Human
,—__—-,
–,
Factors
inComputiig
Systems .

likely consequences for proposed variants of these systems themselves logarithmically, when the user’s shift among
in terms of the metric. Finally, we introduced the next step available access methods is taken into account. This would
in the development of this analysis, taking into account the be interesting to know as well as knowing what are
frequency of access for different items of information, characteristic values for the exponents.

While we have explored direct-walk information access ACKNOWLEDGEMENTS


because of its simplicity, other more complex information The Spiral Calendar was designed and implemented in the
access dialogues should be able to be analyzed with this Information Visualizer by Robert DeLine (University of
basic method and it should be possible to elaborate the Virginia and Carnegie-Mellon University) as a summer
analysis for more insight. project at Xerox PARC. An improved version has now
been integrated into the standard Information Visualizer
The purpose of the paper was to move some of the release by George Robertson (Xerox PARC).
qualitative reasoning behind the design of the Information
Visualizer and other information access systems closer to a REFERENCES
measurable, computable methodology. In this way, we 1. Anderson, J, R, and Schooler, L. J. Reflections of the
hope to be able to understand more precisely the environment in memory. Psychological Science 2(6
consequences of design decisions in this area. Indeed, in November), 1991:396-408.
this case, we discovered the initial design of the Spiral
2. Card, S. K., Moran, T. P., and Newell, A. The
Calendar, while good for handling very large time periods,
Psychology of Human-Computer Interaction. Hillsdate,
was under-optimized for the frequent close-to-present dates.
New Jersey: Erlbaturt, 1983.
While obvious to the designers after the measurement, it
was previously under-appreciated. A number of design 3. Card, S. K., Robertson, G. G., and Mackinlay, J. D.
suggestions have ensued, many couched in terms of what The Information Visualizer, an information workspace.
was required to move various pieces of the curves in Fig. 5. In Proceedings of CHI ’91 ACM Conference on Human
Factors in Computing Systems (New Orleans,
It should be noted that individual differences, as indicated by
Louisiana, April 27–May 2, 1991). ACM, New York,
the standard deviation in Tables 1 and 2, are relatively
1991, pp. 181-188.
small, suggesting the results are not very sensitive to
individual users. The method is therefore usable by an 4. Henderson, D. A., Jr. and Card, S. K. Rooms: The use
individual system budder with a stop watch, timing himself of multiple virtual workspaces to reduce space
or herself. Indeed, a pilot of this experiment with one of contention in a window-based graphical user interface.
the authors timing himself with a stopwatch while ACM Transactions on Graphics 5 (3, July 1986).,21 1-
performing the role of the user yielded a very similar curve 243.
to that plotted in Fig. 5 (with 5 s/cycle instead of 3.5
s/cycle). This is a significant finding, because, while it is 5. Mackinlay, J. M. and DeLine, R. Designing calendar
important to do formal user testing in building systems, it visualizes for the Information Visualizer. Research
is also important to have inexpensive methods designers Report, Xerox PARC, Palo Alto.
can use rapidly as they work to reduce the number of more
expensive user tests required. This is similar to the way 6. Robertson, G. G., Card, S. K., and Mackinlay, J. D.
many experienced system builders now routinely perform Information visualization using 3D interactive
animation. Communications of the ACM, 36 (4,
system timings as they work.
April), 1993,57-71.
Finally, according to Fig. 5, the number of information
7, Russell, D. M., Stefik, M, J., Pirolli, P., and Card, S.
items accessible by these systems increases logarithmically
K. The cost structure of sensemaking. In Proceedings of
with time cost, roughly at
CHI ’93, ACM Conference on Human Factors in
Software (April 24-29, Amsterdam). New York ACM,
Number of item accessible= A eO’5 t ,
1993, pp. 269-276,
where t is the cost of access in seconds. This is as
8. Tennant, H. and Heilmeier, G. H. Knowledge and
expected, since the user has a sttccession of choices, each
equality: Harnessing the tides of information abundance.
with a similar branching factor. But the cost structure for
In Leebaert, D. (cd.), Technology 2001: The Future of
other information systems may also tend to arrange
Computing and Communications. Cambridge,
Massachusetts: The MIT Press, 1991,

244

You might also like