Xerox Palo Alto Research Center
3333 Coyote Hill Road
Palo Alto, CA 94304
pirolli@parc.xerox.com
card@parc.xerox.com
To motivate previous designs of interactive information
systems
[6] , we
appealed to mechanisms of cognitive science [5], and to
general principles of information science
[17] .
We have argued that in an information-rich world, the real
design problem to be solved is not so much how to collect
more information, but rather, how to optimize the user's
time, and we have deployed these principles in an attempt to
increase relevant information gained per unit time expended.
But for task analysis, design exploitation, and evaluation of
information systems, a more developed theory is needed.
In this paper, we lay out the framework for an approach we
call information foraging theory. This approach considers
the adaptiveness of human-system designs in the context of
the information ecologies in which tasks are performed.
Typically, this involves understanding the variations in
activity afforded by some space of human-system design
parameters, and understanding how these variations trade-off
the value of information gained against the costs of
performing the activity. While complementary with
information processing approaches to computer interaction
theory, such as those in the GOMS family [5, 14],
information foraging theory emphasizes a larger time-scale
of behavior, the cost structure of external information-
bearing environments, and human adaptation.
Consider the time-scales of activity outlined by Newell
[13] .
The sorts of information-seeking and sensemaking activities
of interest to us span from the middle of the cognitive band
of activity (~100 ms - 10 s), across the rational band
(minutes to hours), and perhaps into the social band (days to
months). Typically, information-processing models of
cognition have addressed behavior at the cognitive band,
and elementary cognitive mechanisms and processes (e.g.,
such as those summarized in the Model Human Processor,
[5] )
play a large part in shaping observed behavior at that grain
size. As the time scale of activity increases, "there will be a
shift towards characterizing a system...without regard to the
way in which the internal processing accomplishes the
linking of action to goals" (Newell [13], p 150). Such
explanations assume that behavior is governed by rational
principles and largely shaped by the constraints and
affordances of the task environment. Rather than assuming
classical normative rationality, one may assume that the
rationale for behavior is its adaptive fit to its external
ecology [4] .
This is the essence of an ecological stance (Neisser, as cited
in Bechtel [4]) towards cognition. Whereas information-
processing models, such as GOMS, provide mechanistic
accounts of how cognition operates, ecological models
address why it operates that way, given the ecological
context in which it occurs. This kind of integrated
explanatory framework has been promoted by Marr
[12] and, more recently, Anderson [1, 2] in cognitive science.
Optimal foraging theory is a theory that has been developed
within biology for understanding the opportunities and
forces of adaptation. We believe elements of this theory can
help in understanding existing human adaptations for gaining
and making sense out of information. It can also help in task
analysis for understanding how to create new interactive
information system designs.
Optimality models in general include the following three
major components.
We assume that information foraging is usually a task that is
embedded in the context of some other task and its value and
cost structure is consequently defined in relation to the
embedding task and often changes dynamically over time [3,
18]. The value of information
[16]
and the relevance of specific sources
[18]
are not
intrinsic properties of information-bearing representations
(e.g., documents) but can only be assessed in relation to the
embedding task environment.
Usually, the embedding task is some ill-structured problem
for which additional knowledge is needed in order to better
define goals, available courses of action, heuristics, and so
on [15, 20]. Such tasks might include such things as
choosing a good graduate school, developing a financial plan
for retirement, developing a successful business strategy, or
writing an acceptable scientific paper. The structure of
processing and the ultimate solution are, in large part, a
reflection of the particular knowledge used to structure the
problem space. Consequently, the value of the external
information may often reside in the improvements to the
outcomes of the embedding task.
The use of optimality models should not be taken as a
hypothesis that human behavior is classically rational, with
perfect information and infinite computational resources. A
more successful hypothesis about humans is that they exhibit
bounded rationality or make choices based on satisficing
[19]. However, satisficing can often be characterized as
localized optimization (e.g., hill-climbing) with resource
bounds and imperfect information as included constraints
[23]. Optimality models do not imply that animals or
information foragers will necessarily develop so as to
embrace the simple optimum. Rather, they describe the
possibilities of a niche, a possible advantageous adaptation if
not blocked by other forces (for example, the consequences
of another adaptation). For us, these models help fill in
what Anderson [1] calls the Rational Level theory of
information access.
In the context of a an analysis of the Scatter/Gather
document browser [9] we introduce two simple models, the
information patch model, and the information diet model,
borrowed rather directly from optimal foraging theory.
These "conventional" models derive from Holling's disc
equation
[22], which
states that rate of currency intake, R, is the ratio of net
amount of currency gained (energy in the case of biological
systems; information value in our case), U, divided by the
total amount of time spent searching, Ts, and exploiting, Th,
TABLE OF EQUATIONS. Equation 1
TABLE OF EQUATIONS. Equation 2
We also develop a more comprehensive dynamic model that
incorporates the information patch model and information
diet model as subcomponents. Using dynamic programming
we illustrate how one may determine the optimal human-
system strategies using dynamic programming techniques.
We discuss an information patch model in the context of the
Scatter/Gather text database browser.
Conceptually, a collection may be clustered, for instance,
into B = 10 groups of related documents. Each cluster is
represented by a separate area as in Figure 1. For each
cluster, the user is presented with typical words that occur in
the text contents of documents in a cluster, as well as the
titles of the three most prototypical documents. The user
may gather some subset of those clusters, by pointing and
selecting buttons above each cluster, and then ask the system
to scatter that subcollection into another B subgroups, by
selecting a Scatter/Gather button at the top of the display in
Figure 1. The clustering is based on a form of inter-
document similarity computation based on representations of
text contents. Scatter/Gather browsing and clustering
employs methods that can occur in constant interaction-time
[8].
FIGURE 1. The Scatter/Gather interface for navigation
through large document collections.
This simple loop of activity can be characterized by a
cumulative gain function gi(t) that indicates how much
information value is acquired over time t in cluster-patches
of type i. In our empirical studies, we used specific task
instructions that indicated that the information value was
simply the number of relevant documents collected. The
proportion of relevant documents in a cluster is the
precision, P, of that cluster
Abstract
Information foraging theory is an approach to the analysis of
human activities involving information access technologies.
The theory derives from optimal foraging theory in biology
and anthropology, which analyzes the adaptive value of
food-foraging strategies. Information foraging theory
analyzes trade-offs in the value of information gained against
the costs of performing activity in human-computer
interaction tasks. The theory is illustrated by application to
information-seeking tasks involving a Scatter/Gather
interface, which presents users with a navigable,
automatically computed, overview of the contents of a
document collection arranged as a cluster hierarchy.
Keywords:
Information foraging theory, information
access.
Introduction
Recent years have witnessed an explosive increase in public
interest in information access and communication
technologies. Along with this burgeoning excitement, the
rapid growth of electronically available information sources,
such as those available on the Internet, has further
exacerbated the need for effective and efficient tools for
information workers and consumers. For researchers and
developers in human-computer interaction, this increases the
need for models and analysis techniques that allow us to
determine the value added by particular information access,
manipulation, and presentation techniques, and to reveal
design elements that may yield further enhancements.
INFORMATION FORAGING TASKS
Information foraging refers to activities associated with
assessing, seeking, and handling information sources. Such
search will be adaptive to the extent that it makes optimal
use of knowledge about expected information value and
expected costs of accessing and extracting the relevant
information. We use the term "foraging" both to conjure up
the metaphor of organisms browsing for sustenance and to
indicate a connection to the more technical optimal foraging
theory found in biology and anthropology [21, 22].
Animals adapt their behavior and their structure through
evolution to survive and reproduce to their circumstance.
Essentially animals adapt, among other reasons, to increase
their rate of energy intake. To do this they evolve different
methods: a wolf hunts ("forages") for prey, but a spider
builds a web and allows the prey to come to it. Humans
seeking information also adopt different strategies,
sometimes with striking parallels to those of animal foragers.
The wolf-prey strategy bears some resemblance to classic
information retrieval, and the spider-web strategy is like
information filtering. Human hunter-foragers have been
observed to hunt in groups when the variance of finding food
is high. They accept a lower expected mean in order to
minimize the probability of several days without food.
Similarly, we have observed, in the field, professional
market analysts who had developed an ethic of cross-
referring information, essentially information-foraging in
groups, so as to reduce the probability of missing important
literature.
OVERVIEW OF THE EXAMPLES
We present several examples of foraging analyses to
illustrate some of the range of problems and insights that
may be addressed. Our coverage has to be limited, so we use
three relatively concrete and detailed analyses from a
particular system.
The information patch model and the information diet model
are formulated as variants of Equation 2. We discuss the
analytic optimal solutions to these models in the context of
the illustrations.
EXAMPLE 1: FORAGING IN INFORMATION PATCHES
A user's encounters with valuable or relevant information
will typically have a clumpy structure over space and time.
Information items are often grouped into collections such as
libraries, databases, and wire services. The biological
analogy is that an organism's ecology may have a variety of
food patches of differing characteristics and the organism
must decide how to best allocate its foraging time. Models
of this situation are called patch models [7, 10,
22] .
Scatter/Gather Interface
Figure 1 presents a typical view of the Scatter/Gather
interface. [Footnote1].
The emphasis in this browsing paradigm is to
present users with a kind of automatically computed
overview of the contents of a document collection, and to
provide a method for navigating through this summary at
different levels of granularity. This is achieved by
organizing the collection into a cluster hierarchy.
Patch Analysis
We can view Scatter/Gather clusters as information patches.
Foraging in a cluster-patch corresponds to selecting a cluster,
displaying the document titles belonging to the cluster in a
scrollable window, scanning/scrolling through the listed
titles, and for each title deciding if it is relevant or not to the
query at hand. If the title is judged relevant, then it is
handled by selecting, cutting, and pasting it to a query record
window. Relevant and irrelevant documents will be
randomly intermixed in the display window.
TABLE OF EQUATIONS. Equation 3
,
.
, 
that maximizes the
currency intake rate R.
