Conference PaperPDF Available

Visualizing and Assessing Navigation in Hypertext.


Abstract and Figures

User navigation has been a central theme in both theoretical and empirical work since the earliest days of hypertext research and development. Studies exploring user navigation have, however, tended to rely on indirect navigational measures and have rarely tried to relate navigation to performance solving problems or locating information. The purpose of this paper is to propose methods that lead to a more direct representation and analysis of user movement in hypertext and to empirically explore the relationship of resulting measures to performance in a hypertext search task. Results of this study support the claim that the proposed graphical and numerical methods have empirical significance and may be useful in applications related to assessing and modeling user navigation.
Content may be subject to copyright.
Visualizing and Assessing Navigation in Hypertext
John E. McEneaney
Division of Education
Indiana University South Bend
South Bend, IN 46634-7111, USA
Tel/Fax: 219-237-4576 / 219-237-4550
User navigation has been a central theme in both theoretical
and empirical work since the earliest days of hypertext
research and development. Studies exploring user
navigation have, however, tended to rely on indirect
navigational measures and have rarely tried to relate
navigation to performance solving problems or locating
information. The purpose of this paper is to propose
methods that lead to a more direct representation and
analysis of user movement in hypertext and to empirically
explore the relationship of resulting measures to
performance in a hypertext search task. Results of this study
support the claim that the proposed graphical and numerical
methods have empirical significance and may be useful in
applications related to assessing and modeling user
KEYWORDS: visualization; user paths; path analysis;
navigation patterns; navigation metrics; empirical
In response to problems related to hypertext navigation,
researchers and developers have created a variety of
powerful tools, many of them based on visualization
techniques. Site maps are now commonly used and there is
evidence that users find them helpful in navigating and in
establishing a clearer idea of the organizational structure of
a site [8, 32]. In larger networks where complete site maps
are impractical, fish-eye views [12, 28, 2], clustering
techniques that organize nodes into meaningful groups [13,
21], and a variety of other filtering and mapping techniques
(e.g. [15, 22]) have been developed to assist both in
development and use of large scale hypertext networks.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distrinbuted for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists requires prior specific
permission and/or a fee.
Hypertext 99 Darmstadt Germany
Copyright ACM 1999 1-58113-064-3/99/2...$5.00
Numerical measures and metrics have also been
proposed to help assess global properties of hypertext that
may be related to the difficulties users experience. Some of
these metrics [1] have been imported directly from other
disciplines, most notably from social network theory, where
there is a long history of using proximity measures that have
been more recently applied in hypertext research. Other
metrics have been developed to address the specific needs
and perspectives of hypertext researchers (e.g. [3]).
There has also been interest in developing a better
understanding of how users navigate hypertext under
various task and environmental conditions. Some of these
studies have employed “static” measures related to numbers
of nodes or links accessed, measures of time and path length
[29, 26, 19] or analyzed selected episodes of movement,
tabulating navigation within or across sections of a network
[29]. Some investigators have applied statistical techniques
to identify “clusters” of nodes and interpret user navigation
in terms of these constructs [17]. Other statistical
approaches include collapsing data from large numbers of
users into state transition probability tables that are used as
a basis for analysis [7], and identification of statistical
benchmarks that might be useful in developing theoretical
models [26, 23].
Perhaps the most consistent trend in user navigation
research however is the use of navigational “paths” as a
primary source of data. Although the concept of a path has
been referred to using a variety of terms including “route”
[4], “user trail” [14], “audit trail” [20] and others, these
terms all refer to data sets that record the sequence of nodes
visited by a subject in a hypertext session and often also
include measures of time related to visits. That user paths
are commonly employed in navigational studies should be
no surprise. Although a time-stamped path misses
deliberations that go into a users’ decision making (e.g. a
pointer hovers momentarily over one link before moving on
to another that is clicked), a path represents the single most
complete measure of user navigation and thus affords an
important window on the search process [17]. Moreover,
since data can be recorded and formatted in an unobtrusive
manner on-the-fly, this approach provides empirical
investigators with a powerful data collection tool that has
the additional benefit of being seamlessly integrated with the
delivery of experimental materials.
For all of the interest in navigational paths however
there have been relatively few studies that have sought to
examine the relationship between patterns of navigation and
outcome measures. Two exceptions are studies by Smith
[30] and Cardle [5], both of whom investigated the
relationship of informal measures of search success (i.e.
displays of frustration or confidence by subjects) with a path
efficiency measure. Both of these studies, however, were
based on static navigational measures (e.g. number of pages
visited) that did not attempt to incorporate spatial or
directional aspects of user paths. Other studies that sought
to relate path data to outcome measures have been based on
similarly indirect path measures. Chang and McDaniel [6],
relied on video transcripts of their subjects and
accompanying “think aloud” data to subjectively categorize
navigational patterns. Pirolli, Pitkow, and Rao [25] and
Chen [7] adopted a more quantitative focus on user paths
but employed large-scale aggregate data drawn from web
server statistics. By using these aggregated data sets,
however, the “paths” studied had to be inferred from access
logs and were thus subject to problems related to firewalls,
proxy masking of user identity, intentional reloading of
documents by users, and missed hits as a result of local
browser caches.
Although discrete static measures, informal descriptive
characterizations, and aggregated summarizations of user
navigation are useful starting points, our understanding of
user navigation will be significantly enhanced if user
movement can be more directly represented, analyzed, and
related to performance outcomes. It is the purpose of this
paper to achieve these ends - to define methods that support
more direct representation and analysis of user paths and to
empirically relate resulting measures to hypertext search
In this paper, user paths will be analyzed in two
different ways. One form of analysis is based on a graphical
method intended to illustrate user paths in a way that makes
navigational patterns visually distinct. The second form of
analysis relies on path-specific structural metrics that are
related to the visually distinct categories identified by
graphic analysis. Following definition of these path metrics,
the paper will report on an experimental study that explores
the relationship between the proposed measures and a
quantitative search outcome measure. The paper will
conclude by considering limitations of the proposed methods
and will suggest some potential applications of the proposed
methods and metrics.
The framework adopted to visualize and assess user
navigation is based on the traditional node-and-link model.
At least part of the popularity of this model can be attributed
to two simple but powerful formalisms that support analysis:
adjacency matrices that are well suited to computational
analysis and labeled directed graphs (digraphs) that present
structural information in a readily interpreted visual format.
Both of these formalisms have played important roles in
earlier work establishing metrics to assess structure in
hypertext and they serve similar roles in the present study.
The two structural metrics of special interest in the
present study are compactness and stratum [3]. The purpose
of these metrics is to yield global network-based assessments
of structure that are grounded in node-based centrality
measures. Briefly, compactness refers to the overall
connectedness of a network with more sparsely linked
networks resulting in values for compactness close to 0,
while densely connected networks yield compactness closer
to 1 (a continuum labeled “sparse vs. rich” hypertext by
Nielsen [24]). Stratum, on the other hand, refers to the
degree of linearity of a network, as indicated by the extent to
which a network is organized so that certain nodes must be
read before others. Stratum also ranges between 0 and 1,
with more linear networks closer to 1 and more web-like
networks closer to 0.
Although compactness and stratum were originally
developed to assess the structure of hypertext networks,
subsequent work [27] has suggested that these same metrics
might be usefully applied as tools in assisting users navigate
networks. The present study attempts to broaden the
application of these metrics still further by proposing
adaptations intended to support the analysis of user
movement in hypertext. Before considering how these
metrics can be adapted, however, it will be useful to be more
specific both about the larger conceptual framework that
supports the proposed methods and the concept of a path.
Conceptual Preliminaries
Distance matrices have proven to be particularly useful
for analyzing hypertext networks. Each cell Dij in a distance
matrix identifies the minimum number of steps required to
move from node i to node j. By convention, the distance
from a node to itself is “0” and the distance to unreachable
nodes is infinite. In circumstances where infinite values
complicate subsequent computation, they are replaced with
a conversion constant K (usually set equal to the number of
nodes in the network), resulting in a converted distance
matrix. Sums across rows in the converted distance matrix
(3jCij) are referred to as converted out-distances and
represent the centrality of nodes when considered as points
of departure. Sums down columns in the distance matrix
(3iCij) are referred to as converted in-distances and
represent the centrality of nodes as destinations. The sum
of all entries in the converted matrix is referred to as the
converted distance (CD) of a network and is commonly
A. Path:¢¢6,30,6,37,6,15,16,6,21,6,21,30,6,23,24,6,35,6¦¦.
from 6 15 16 21 23 24 30 35 37
6 010210111
15 001000000
16 100000000
21 100000100
23 000001000
24 100000000
30 200000000
35 100000000
37 100000000
B. Path matrix for the path above.
Figure 1. A path across a larger 78-node
hypertext (A) and its resulting path matrix. (B).
from 6 15 16 21 23 24 30 35 37
6 012112111
15 201334333
16 120223222
21 123023122
23 234301333
24 123220222
30 123223022
35 123223202
37 123223220
A. Converted distance matrix for the path in Figure 1A.
B. Path diagram for the path in Figure 1A.
Figure 2. The converted distance matrix and
path diagram for the path in Figure 1A.
employed to normalize in- and out-distances so that
comparisons can be made between networks that differ in
size or connectedness.
A path P in a network is simply the sequence of nodes
or pages visited by a reader during a browsing session (i.e.
P = +p1,p2,p3,...,pL,, where L refers to the length of the path
or number of node visits). Although this use of the term
“path” deviates from its meaning in the underlying graph-
theoretic framework (since it allows nodes to appear more
than once), this usage has been adopted in recent work [25,
10, 9] and will be employed in the present study as well.
Since nodes can occur in more than one position in a path,
however, the length of a path may not correspond to the
number of distinct nodes visited and this has some important
consequences for both path visualization and the calculation
of metrics based on path matrices.
The Path Matrix And Its Derivatives
Unlike the distance matrix described earlier that
represents distances between nodes, a path matrix represents
frequencies of node transitions during a browsing session
from each node in the path to every other node in the path
[25]. If the hypertext under consideration is closed, the path
matrix can also be normalized by representing every node in
the hypertext as a whole, regardless of whether a node
appears in the path. As a result of this expansion, it is
possible to sum individual user paths into group “paths”.
The normalizing expansion of the path matrix is achieved
by inserting rows and columns filled with zeroes in the
appropriate places in the path matrix so that every node in
the hypertext is represented. The net result of this
expansion is to embed the path within the larger structure of
the hypertext. The structural features of the original path
matrix are, however, preserved while establishing a normal
form for all paths.
Calculation of the proposed path metrics, like their
structural counterparts, requires a distance matrix and
suitable conversions. Consider, for example, the path in
Figure 1A, consisting of 17 transitions, beginning and
ending at node 6. The path matrix is constructed by
creating a suitably labeled matrix that includes each distinct
node in the path and then incrementing the appropriate cell
for each transition represented. The resulting path matrix
(Figure 1B) indicates the number of transitions from each
node to every other node in the path.
The distance matrix and converted distance matrix for
a path are created through straightforward adaptations of the
procedures described by Botafogo et al. [3, 27], with one
important difference. Creation of a distance matrix for a
path begins by substituting a value of “1” for all entries in
the original path matrix that exceed one (i.e., that represent
multiple transitions) and then generating a distance matrix
as would be done with an adjacency matrix representing a
hypertext. For path stratum, no further conversion is
needed. The path compactness metric, however, requires
conversion, replacing infinite cell entries with the
conversion constant K, where K equals the number of nodes
in the original path matrix (Figure 2A). The path diagram
is constructed from the original path matrix by creating a
vertex for each node represented and an arc for each non-
zero cell entry (Figure 2B).
Path Metrics
Although the computational procedures involved in
generating path metrics are straightforward manipulations
of the structural metrics they are based on [3, 27], there is an
important difference between path matrices and
corresponding structural matrices. In structural matrices,
every node in the hypertext is represented but this is clearly
not the case in the path matrix presented in Figure 1B,
which only represents nodes actually in the path. While the
expanded path matrix does represent all nodes in the
hypertext it will usually be much larger than the path matrix
and this has significant consequences for both path
visualization and metrics, suggesting that the path matrix
rather than the expanded path matrix should serve as the
basis for subsequent visualization and calculation.
With respect to visualization, use of the expanded path
matrix will tend to crowd the display of a path with large
numbers of isolated (i.e. unconnected) nodes that do little
more than communicate the size of the network traversed,
something more easily done by simply reporting this
information. A second problem introduced by using the
expanded path matrix is that subsequent calculations will
tend to be dominated by the conversion constant. Since
most paths traverse only small portions of hypertexts, the
expanded distance matrix will typically include large
numbers of unreachable nodes with infinite entries that, on
conversion, will be replaced by K, the conversion constant.
Although the influence of the conversion constant can be
moderated somewhat by scaling this constant up or down, it
seems counterproductive to define a metric in a way that
knowingly obscures the variability that is of greatest interest.
Although a complete representation of nodes is clearly
needed in assessing network structure, trying to distinguish
user paths within a visual display of a larger network seems
ill-advised. The very nature of path analysis, particularly
when it attempts to treat individual users, suggests that
ignoring unvisited nodes is a better approach since any
attempts to account for the influence of these unvisited
nodes is speculative at best. Moreover, if as cognitive
flexibility theory suggests [31, 16], users define their
“knowledge space” as a result of their particular experiences
traversing larger spaces, it seems theoretically justifiable (as
well as computationally convenient) to base path
visualization and metrics on the smaller path matrix.
As a consequence of these considerations, it is possible
to define path compactness and path stratum, the specialized
metrics that are a central objective of this study. Path
compactness (PCp) refers to the complexity of a user’s path,
based on the same notion of connectedness employed in the
corresponding structural metric and is formally defined as
PCp '(PMax&EiEjPCij )
where PC refers to the converted distance matrix of the path
and PMax and PMin refer, respectively, to the maximum and
minimum converted distance values that the path matrix can
assume for a completely-connected (PMin) and completely-
disconnected (PMax) network consisting of as many nodes as
are in the path. PMax and PMin are given by
PMax'K(n2&n),and PMin'(n2&n),
where n is simply the order of the path matrix.
Path stratum (PSt) is also defined in a manner analogous
to its structural equivalent with
PSt 'path absolute prestige
where the absolute prestige of a path is defined similarly to
a network with the exception that the distance matrix is
derived from a path matrix rather than an adjacency matrix.
The linear absolute prestige (LAP) that serves as the
normalizing measure for this metric is also defined
analogously to the structural metric, based on the number of
distinct nodes (n) in the path with
4,if nis even.
4,if nis odd.
The purpose of this section is to describe the results of
an empirical study designed to assess whether the proposed
measures are empirically meaningful in the sense that they
can be shown to be related to objective categories of users.
More specifically, the focus of the present validation will be
to determine whether the proposed graphic techniques and
path metrics can be shown to be associated with user success
in a hypertext search task.
Participants in the validation study were adult students
at a medium-sized Midwestern public university in the USA.
Identities of subjects were coded so that individuals could
not be identified and all procedures were reviewed and
approved by a university human subjects board. A total of
89 teacher education students participated, with data
collection extending across both terms of the 1997-1998
academic year. The experiment required subjects to respond
to a set of academic advising questions using an electronic
student advising handbook. Subjects were to answer as
many questions as possible within a 15-minute period. The
handbook consisted of approximately 31,000 words in 78
text nodes structured in a hierarchical-linear fashion with
major handbook divisions organized hierarchically and
nodes within those divisions organized in a linear fashion.
The handbook duplicated the content and overall structure
of a print version that had been in use for a number of
Since all subjects participating in the study were new
Figure 3: Three path diagrams for
subjects whose scores on the hypertext
search task were low.
Figure 4. Three path diagrams for subjects
whose scores on the hypertext task were high.
admissions to the teacher education program and had not yet
participated in formal academic advising, it was unlikely
they were familiar with the handbook content. Moreover,
since all subjects were at the same point in their academic
careers, it is highly unlikely there was any systematic
variation of subjects’ familiarity with program policies and
procedures across the experimental groups.
Following completion of the experimental sessions,
subjects’ browser cookie files were retrieved and path data
were extracted. Mathematica [33] routines were developed
to format graph files for display by GraphViz 1.3 [11],
employing a hierarchical embedding format. Mathematica
routines were also used to calculate the path compactness
and path stratum metrics described above. Subjects’
responses to academic advising questions were scored on the
basis of information provided in the handbook across a scale
of three values with 0 points awarded for incorrect and
omitted responses, ½ point for partially correct and correct-
but-incomplete responses, and 1 point for complete and
correct responses [30, 18, 34].
Visual Analysis
The first stage of analysis validating the proposed
graphical methods was based on a subset (n = 29) of the
larger subject pool (n=89). In this preliminary graphical
analysis subjects were grouped according to their search
performance answering questions using the hypertext
handbook. The “high” group consisted of subjects with the
top three scores in each of four groups that had been set up
to counter-balance experimental conditions. The “low”
group, on the other hand, consisted of subjects with the
bottom three scores in each of the four counter-balancing
groups. As a result, the analysis was based on two groups of
12 subjects that differed according to their success in
carrying out the hypertext search task.
The next step in the analysis was to create and review
path diagrams for the subjects in the two groups, with the
intent of discerning visually distinctive patterns that might
be related to success in the search task. In reviewing these
path diagrams, it soon became apparent that there were,
indeed, visually salient features that seemed to be related to
subjects’ success in the experimental task. A number of
examples of these distinctive path diagrams for individual
high and low scoring subjects are presented in Figures 3 and
Review of the path diagrams suggested that subjects
whose scores on the search task were low tended to assume
a “passive” approach to locating answers in the handbook,
relying much more heavily on sequential “page-turning”
than those subjects who did well on the search task. The
path diagrams for these low scoring subjects typically
revealed distinctively linear patterns of movement. Path
diagrams for high scoring subjects, on the other hand,
tended to show shallow hierarchical patterns of movement
with the handbook table of contents serving as the root of
the tree, indicating repeated visits to the table of contents
during the course of the browsing session. Moreover, as
indicated in Figures 5 and 6, similar navigational patterns
resulted when group paths were generated by summing the
individual expanded path matrices for high- and low-scoring
Since some link traversals in group paths probably
represent idiosyncratic thinking and navigational errors,
“noise” was eliminated from the summed group diagrams by
Figure 5: Group path matrix (33pi ) for
low ability hypertext readers for path
transition frequencies $$ 3.
Figure 6. Group path matrix ( 33pi ) for high ability
hypertext readers for path transition frequencies
$$ 3.
setting a threshold that had to be met in order for a
traversal to be displayed. In figures 5 and 6, the threshold
is set equal to 3, with the result that only those links are
displayed that were traversed at least three times by the
subjects in each group. While setting thresholds to other
values (5,2,1, etc.) altered diagrams in minor ways (mainly
by increasing the number of links displayed), alternative
settings did not alter the characteristic linear and shallow
hierarchical patterns associated with the low- and high-
scoring groups.
The graphical analysis carried out provides fairly
compelling, if informal, evidence in support of distinctive
navigational patterns associated with hypertext search
outcome measures. More effective hypertext search is
associated with a shallow hierarchical path diagram that
results from subjects making repeated trips back to the main
table of contents in order to make decisions about how to
locate information in the electronic handbook. Less
effective hypertext search is associated with a more passive
linear path diagram that reflects users’ reliance on
sequential “page-turning” with users hoping to locate
desired information by simply coming across it in their
Moreover, these graphical analyses are suggestive about
what we might expect to find when we examine the
association between the path metrics that have been defined
and hypertext search outcome scores. Specifically, the
linear character of the low scoring group suggests that path
stratum will be negatively correlated with a subjects’
hypertext search score since low scoring subjects seemed
more likely to adopt linear navigational paths. Conversely,
since the compactness of a bidirectional star pattern
(characteristic of the paths of more successful subjects) tends
to approach 1 as the number of nodes increases, while both
bidirectional cycles and linear patterns (characteristic of less
successful subjects) approach values less than 1 [3], it
appears likely that path compactness will be positively
associated with users’ hypertext search scores. The research
hypotheses that will be explored in the quantitative analysis
that follows will therefore involve one-tailed, directional
tests of significance of the following research hypotheses:
1) The path compactness metric will correlate significantly,
in a positive fashion, with subjects’ hypertext search
scores, and
2) The path stratum metric will correlate significantly, in an
inverse fashion, with subjects’ hypertext search scores.
Analysis Based on Path Metrics
The purpose of this section is to present results of
analyses carried out to determine if the predictions made on
the basis of the graphical analysis hold up under quantitative
analysis. As noted earlier, subjects’ outcome measures are
based on numbers of questions answered correctly using a
hypertext student advising handbook. Pearson correlation
coefficients were determined relating a number of
experimental variables and hypertext search scores,
including path compactness and path stratum. Results of
the analyses are indicated in Table 1, with significant
correlations (" < .05) flagged. Note that, of the variables
examined, only path compactness and path stratum
correlated significantly with the hypertext search measure.
Correlations of Various Experimental Variables
with Hypertext Search Scores
r p n
Print Ability .210 .137 29
Pages viewed -.006 .955 89
Order of path matrix -.011 .921 89
Path compactness .239* .012† 89
Path stratum -.205* .027† 89
* p < .05; † Indicates a one-tailed test of association. All other
tests are two-tailed.
Table 1. Correlations (with p values and
subjects) for experimental variables and the
hypertext search scores.
Moreover, these correlations were as expected, with
compactness exhibiting a significant positive correlation and
stratum exhibiting a significant inverse correlation. These
analyses suggest that the observed relationship is not likely
to be the result of chance, and thus support the interpretation
of path diagrams and the proposed metrics as reflecting
empirically meaningful and potentially useful measures of
hypertext navigation.
Results of the empirical validation suggest that the
proposed methods and metrics can be productively applied
in assessing user navigation. Moreover, the results reported
suggest that navigational patterns and their associated
metrics may be useful as indirect measures of user strategy
and perhaps even of users’ success in cognitively
“modeling” the domain represented by a hypertext. If, as
cognitive flexibility theory suggests, learning in hypertext
materials involves the cognitive reconstruction of a domain
space through repeated traversals across that space, the
paths users choose are sure to have a powerful influence on
learning outcomes. In the present study, subjects who
adopted shallow, hierarchical search strategies that more
accurately “modeled” the organization of the hypertext
materials were more successful in their search, while those
who adopted more linear paths through the materials were
less successful. In effect, more successful subjects
recognized and took advantage of the structure of the
domain space by returning to the higher ground of the table
of contents and the broader cognitive view it afforded of the
That more successful hypertext users recognize and take
advantage of the structure of the materials they are using is
not, in itself, very surprising. That is, after all, the purpose
of graphic overviews, site maps, and other techniques that
have been shown to promote more effective use of hypertext.
What is important about the methods and metrics proposed
is that they are not merely ad hoc constructions, but are
grounded in a widely used conceptual framework, and that
they hold up under both qualitative and quantitative
scrutiny. Although prior work has identified qualitative
features similar to those noted here, informal
characterizations such as “loopiness” and “spikiness” [34]
can now be related to objectively assessed metrics. Another
important feature of the methods proposed is that they are
based on information that is immediately and unobtrusively
available during reading, something that is not generally
true of outcome measures. Given the demonstrated
association of path information and outcome measures, it
may be possible to apply real-time path data in generating
user models that will lead to more effective adaptive
hypertext systems. It may also be possible to apply these
metrics in designing user paths to meet particular objectives
or needs of users. Even in the absence of immediate
applications, however, it will be important to explore these
more direct measures of user movement, given the interest
in, and widespread use of, less direct measures.
While the results of the present study are relatively
clear-cut, three limitations suggest that these findings
should be considered preliminary. One limitation has to do
with the strength of the observed association between path
metrics and search success. A second limitation has to do
with the choice that has been made with regard to
normalization, and a third set of related limitations is
associated with the generalizability of findings, given
constraints imposed by the design of the study and the
experimental materials.
Although the observed association between the proposed
metrics and search success is not likely to be due to chance,
the strength of the association is not great. A weak
association remains of significant interest, but it also
suggests that this variable should be considered within a
larger explanatory context. Regarding normalization, it is
relevant to note that Botafogo et al. [3] recognize the general
nature of the normalization problem in their work
establishing structural metrics. In response to this problem,
they suggest that alternative normalization procedures be
considered, particularly for the stratum metric. They note,
for instance, that since stratum depends on LAP, a measure
that is O(n3) for matrices of order n, it may be problematic
to compare networks that have large differences in numbers
of nodes. Given this, the question arises whether the
variation in the order of path matrices across subjects is
sufficient to raise concerns about the analyses that have been
carried out.
Two circumstances of the present study suggest,
however, that the problem of normalization has not
compromised the specific results reported. One
circumstance is that although there was variation in the size
of the path matrices used to calculate path stratum values,
there was no significant correlation between subjects’ search
scores and the size of their associated path matrices (
Pearson r = -.011, p = .921). Path matrices varied, but there
is no evidence of a systematic variation that might influence
the relationship noted between the outcome measure and
path stratum.
The second circumstance is related to the observation
([3], pp. 169-170) that differences in stratum values can
result when index and reference nodes are excluded from the
stratum calculation, the implication being that the presence
of prominent nodes can distort the stratum metric. In an
effort to determine whether this should be a source of
concern in the present study, path stratum values were
recalculated excluding index and reference nodes identified
as those whose in- and out-degrees differed from their
respective means by more than one standard deviation.
Recalculation of path stratum values and reanalysis of the
association between observed path stratum and search scores
remained significant, with only a very minor deviation from
prior results. As noted in Table 1, analysis across all nodes
resulted in an r = -.205. Reanalysis with index and
reference nodes excluded resulted in an r = -.200, suggesting
that the path stratum metric had not been influenced
significantly by unusual user movement across specific
A limitation having to do with the design of the
validation study arises because the investigation has focused
on the relationship between user paths and outcome
measures from a strictly correlational perspective. Had user
movement been more carefully controlled so that a path
“factor” could be established, a stronger inference regarding
the contribution of user movement to search success might
have been possible. That, however, must be addressed in
future research. For the present, we must be satisfied with
the observation that user movement and search success are
significantly associated, without clear understanding about
how and why these measures are related.
Finally, it is important to note that the present
investigation is limited to a test of the proposed methods and
metrics utilizing a single hypertext using a specific browser
interface. The findings reported are consistent with prior
work [4, 34, 35] that suggests hierarchical patterns of
movement benefit readers unfamiliar with the material
presented but larger questions remain. It is, for instance,
still unclear if the proposed methods and metrics will work
equally well in other hypertext structures and under different
browsing conditions. While it is apparent that the observed
efficacy of navigational patterns reflects structural features
of the hypertext itself, which inevitably sets conditions
within which users must operate, it is not clear how specific
hypertext structures influence user movement and the
metrics proposed. Like the limitation related to the design
of the study, this question requires further empirical work.
These limitations notwithstanding, the proposed
methods and metrics afford hypertext developers and
researchers a number of important benefits. One benefit is
that these methods and metrics support more direct analysis
of user movement in hypertext than has been possible
before. A second benefit is that the concepts and
computational framework these methods and metrics are
based on are natural extensions of prior methods and metrics
developed to analyze the structure of hypertext, and thus
support a more general perspective that encompasses both
structure and navigation in hypertext. Finally, there are
both informal and quantitative reasons for confidence in the
adequacy of these methods since the metrics that have been
proposed are clearly related to the graphical displays
developed and these metrics have been shown to have
significant empirical association with success in a hypertext
search task.
This work was supported in part by an Indiana
University Summer Faculty Fellowship. The author would
also like to thank Mr. Michael Mancini for his assistance
collecting data and four anonymous reviewers for
suggestions and comments that contributed to the revision
of this manuscript.
1. Astleitner, H. & Leutner, D. (1996). Applying standard
network analysis to hypermedia systems: Implications
for learning. Journal of Educational Computing
Research, 14(3), 285-303.
2. Bartram, L., Ho, A., Dill, J., & Henigman, F. (1995).
The continuous zoom: A constrained fisheye technique
for viewing and navigating large information spaces.
Proceedings, User Interface and Software Technology
‘95. ACM. New York, 207-215.
3. Botafogo, R. A., Rivlin, E., & Shneiderman, B. (1992).
Structural analysis of hypertexts: Identifying hierarchies
and useful metrics. ACM Transactions on Information
Systems, 10(2), 142-180.
4. Canter, D., Rivers, R., & Storrs, G. (1985).
Characterizing user navigation through complex data
structures. Behaviour and Information Technology,
4(2), 93-102.
5. Cardle, N. T. (1994). A hypercard on Celtic history to
assess navigability measures for hypertext. MSc.
Dissertation. Institute of Information Technology,
University of Nottingham, UK.
6. Chang, C. & McDaniel, E. D. (1995). Informal search
strategies in loosely structured settings. Journal of
Educational Computing Research, 12(1), 95-107.
7. Chen, C. (1997). Structuralizing and visualizing the
WWW by generalised similarity analysis. In Mark
Bernstein, Leslie Carr, & Kasper Østerbye (Eds.),
Proceedings of the Eighth ACM Conference on
Hypertext - Hypertext ‘97, 177-186. New York: ACM.
8. Chen, C. & Rada, R. (1996). Interacting with
hypertext: A meta-analysis of experimental studies.
Human-Computer Interaction, 11, 125-156.
9. Cockburn, A. & Jones, S. (1996). Which way now?
Analysing and easing inadequacies in WWW
navigation. International Journal of Human-Computer
Studies, 45, 105-129.
10. Eklund, J. & Zeiliger, R. (1996). Navigating the web:
Possibilities and practicalities for adaptive navigational
support [On-line]. Available:
11. Ellson, J., Koutsofios, E., & North, S. (1998).
GraphViz 1.3 [Computer software]. Murray hill, NJ:
Lucent Technologies. Available:
12. Furnas, G. (1986). Generalized fisheye views,
Proceedings of CHI ‘86, Human Factors in Computing
Systems, Boston, April, 1986, 16-23.
13. Gloor, P. A. (1991). Cybermap, Yet another way of
navigating in hyperspace. In Proceedings of the Third
ACM Conference on Hypertext, Hypertext ‘91. San
Antonio, Texas, December, 1991, 107-121.
14. Hill, G., Hutchings, G., James, R., Loades, S., Halé, J.,
Hatzopulous, M. (1997). Exploiting serendipity
amongst users to provide support for hypertext
navigation. In Mark Bernstein, Leslie Carr, & Kasper
Østerbye (Eds.), Proceedings of the Eighth ACM
Conference on Hypertext - Hypertext ‘97, 212-213.
New York: ACM.
15. Husemann, H., Petersen, J., Kanty, C., Kochs, H., &
Hase, P. (1997). A user adaptive navigation metaphor
to connect and rate the coherence of terms and complex
objects. In Mark Bernstein, Leslie Carr, & Kasper
Østerbye (Eds.), Proceedings of the Eighth ACM
Conference on Hypertext - Hypertext ‘97, 177-186.
New York: ACM.
16. Jacobson, M. & Spiro, R. (1995). Hypertext learning
environments, cognitive flexibility, and the transfer of
complex knowledge: An empirical investigation.
Journal of Educational Computing Research, 12(4),
17. Lawless, K. A. & Kulikowich, J. M. (1996).
Understanding hypertext navigation through cluster
analysis. Journal of Educational Computing Research,
14(4), 385-399.
18. McKnight, C., Dillon, A., & Richardson, J. (1990). A
comparison of linear and hypertext formats in
information retrieval. In R. McAleese & C. Green
(Eds.), Hypertext: State of the art (pp. 10-19). Oxford:
19. Melara, G. E. (1996). Investigating learning styles on
different hypertext environments: Hierarchical-like and
network-like structures. Journal of Educational
Computing Research, 14(4), 313-328.
20. Misanchuk, E. R. & Schwier, R. A. (1992).
Representing interactive multimedia and hypermedia
audit trails. Journal of Educational Multimedia and
Hypermedia, 1, 355-372.
21. Mukherjea, S., Foley, J. D., & Hudson, S. (1995).
Visualizing complex hypermedia networks through
multiple hierarchical views [On-line].
22. Neves, F. D. (1997). The Aleph: A tool to spatially
represent user knowledge about the WWW docuverse.
In Mark Bernstein, Leslie Carr, & Kasper Østerbye
(Eds.), Proceedings of the Eighth ACM Conference on
Hypertext - Hypertext ‘97, 197-207. New York: ACM.
23. Nielsen, J. (1989). The matters that really matter for
hypertext usability. Proceedings of the ACM
conference on hypertext - Hypertext ‘89, 239-248. New
York: ACM.
24. Nielsen, J. (1990). Hypertext and hypermedia.
London: Academic Press.
25. Pirolli, P., Pitkow, J., & Rao, R. (1996). Silk from a
sow’s ear: Extracting usable structures from the web
[On-line]. Available:
26. Qiu, L. (1994). Frequency distributions of hypertext
path patterns: A pragmatic approach. Information
Processing & management, 30(1), 131-140.
27. Rivlin, E., Botafogo, R., & Shneiderman, B. (1994).
Navigating in hyperspace: Designing a structure-based
toolbox. Communications of the ACM, 37(2), 87-96.
28. Sarker, M. & Brown, M. H. (1994). Graphical fisheye
views, Communications of the ACM, 37, 12, July,
1994, 73-84.
29. Schroeder, E. E., & Grabowski, B. L. (1995). Patterns
of exploration and learning with hypermedia. Journal
of Educational Computing Research, 13(4), 313-335.
30. Smith, P. A. (1996). Towards a practical measure of
hypertext usability. Interacting with Computers, 8(4),
31. Spiro, R. & Jehng, J. (1990). Cognitive flexibility and
hypertext: Theory and technology for the nonlinear and
multidimensional traversal of complex subject matter.
In Don Nix and Rand Spiro (Eds.), Cognition,
Education, Multimedia. Hillsdale, NJ: Lawrence
32. Utting, K. & Yankelovich, N. (1989). Context and
orientation hypermedia networks. ACM Transactions
on Office Information Systems, 7(1), 58-84.
33. Wolfram, S. (1994). Mathematica (Version 2.2)
[Computer software]. New York: Addison Wesley.
34. Leventhal, L. M., Teasley, B. M., Instone, K.,
Rohlman, D. S., & Farhat, J. (1993). Sleuthing in
HyperHolmesTM: an evaluation of using hypertext vs. a
book to answer questions. Behaviour & Information
Technology, 12(3), 149-164.
35. Simpson, A. & McKnight, C. (1990). Navigation in
hypertext: structural cues and mental maps. In R.
McAleese and C. Green (Eds.), Hypertext: State of the
art. Intellect: Oxford.
... Parunak [50] and McEneaney [51] studied the abstract problem of user navigation. Doemel [5] described a system called WebMap, which visualizes the topology of the visited region of the Web graph by calculating the spanning tree. ...
Conference Paper
Full-text available
We research the problem of building knowledge maps of graph-like information. We live in the digital era and similarly to the Earth, the Web is simply too large and its interrelations too complex for anyone to grasp much of it through direct observation. Thus, the problem of applying cartographic principles also to digital landscapes is intriguing. We introduce a mathematical formalism that captures the general notion of map of a graph and enables its development and manipulation in a semi-automated way. We describe an implementation of our formalism on the Web of Linked Data graph and discuss algorithms that efficiently generate and combine (via an algebra) regions and maps. Finally, we discuss examples of knowledge maps built with a tool implementing our framework.
... Parunak [50] and McEneaney [51] studied the abstract problem of user navigation. Doemel [5] described a system called WebMap, which visualizes the topology of the visited region of the Web graph by calculating the spanning tree. ...
We research the problem of building knowledge maps of graph-like information. There exist well-consolidated cartographic principles and techniques for mapping physical landscapes. However, we live in the digital era and similarly to the Earth, the Web is simply too large and its interrelations too complex for anyone to grasp much of it through direct observation. Thus, the problem of applying cartographic principles also to digital landscapes is intriguing. We introduce a mathematical formalism that captures the general notion of map of a graph and enables its development and manipulation in a semi-automated way. We present an implementation of our formalism on the Web of Linked Data graph and discuss algorithms that efficiently generate and combine (via an algebra) regions and maps. We present the MaGe tool, implementing the map framework, and discuss examples of knowledge maps.
... It has been shown that users in a hypertext search task who had hierarchical browsing patterns through the hypertext performed better than users who had sequential browsing paths (McEneaney, 1999). Therefore, it is very important that architectures for electronic handbook allow, or even enforce, such hierarchical patterns, and a concept hierarchy is a good way of doing this. ...
Full-text available
We report on an ongoing project aimed at providing an exemplary architecture for an elec-tronic dissemination environment for scientific handbooks. We focus on our way of facilitat-ing navigation through and access to electronic handbooks by using a WordNet-like concept hierarchy consisting of synsets that are connected to each other and to external sources by semantic relations for navigational purposes.
... A recent information visualization paradigm used to summarize information is that of metro maps (e.g., [12]). Other strands of research related to ours are (visual) navigational histories site maps and bookmarks [9,8,4]. These approaches are designed for human usage and are mainly oriented to visualization not allowing automatic processing, composition and reuse. ...
Full-text available
A map is an abstract visual representation of a region, taken from a given space, usually designed for final human consumption. Traditional cartography focuses on the mapping of Euclidean spaces by using some distance metric. In this paper we aim at mapping the Web space by leveraging its relational nature. We introduce a general mathematical framework for maps and an algebra and discuss the feasibility of maps suitable for interpretation not only by humans but also by machines.
... There have been many implementations of automated user interface operation monitoring, particularly for web browsing, from the usability community. [1,2] In particular Guzdial's recording of user actions in an end-user programming environment, applying labels for five design-state categories and using Markov Chain analysis with his HAWK tool [3] to compare users' transitions between the stages, is reminiscent of our approach though much simpler. There have also been a number of applications to scientific and information visualization, including capabilities for general UNDO and replaying actions, such as [4]. ...
Full-text available
Most visualization taxonomies are designed to help designers and users choose appropriate analysis and visualization techniques for the data being displayed; even automate the generation of visualizations. We have developed a taxonomy with a different purpose: to support the analysis and modeling of the human data exploration process. User actions in a large information visualization and analysis platform are recorded by custom user interface (UI) models that save a history of operations. Each model value is labeled with a taxonomic category. When the global history is viewed as a dataset, these categories provide a standard framework for comparing similar operations throughout the system. Like several other taxonomies, ours is based upon the information visualization pipeline model. However it differs in its focus on user actions, ignoring specifics of data structure. This approach maximizes discovery of common interaction patterns across tools (visualization and analysis) and systems, such as highlighting subsets of the data correlated with certain interactions and allows one to tailor the taxonomy for the specific discovery platform. A key contribution of this work is automating both the recording of actions and their labeling in a principled taxonomy, which greatly reduces the effort required to collect long semantically- labeled histories. Because the approach of taxonomically-labeled UI models that record actions should be usable for any interactive system, it could also facilitate comparing interaction patterns across domains.
... This article defines a theoretical framework that integrates transactional theory (Rosenblatt, 1994a(Rosenblatt, , 1994b(Rosenblatt, , 1995 with recently developed formal methods that support the analysis of reader navigation in hypertext (McEneaney, 1999;2000;2001). Transactional theory provides the larger framework within which formal analysis of reader navigation in hypertext is situated. ...
... The concept of a path has significance for research in hypertext reading from two rather different perspectives. The descriptive perspective uses path data as a basis for empirical investigations of user movement in hypertext (Cardle, 1994;McEneaney, 1999;Smith, 1996). But the concept of path also has a prescriptive interpretation that is more immediately relevant to thinking about hypertext as a learning tool. ...
Full-text available
Abstract One of the fundamental,principles of content literacy instruction is that supporting learning from text should involve a focus both on the content to be learned and on the processes students apply as they work to acquire, organize, and integrate that content (Readence, Bean, & Baldwin, 1995, pp. 9-10). The significance of this fundamental principle in terms of instructionaloutcomes,is clearly related to the accessibility and complexity of the material to be learned, since more complex or less accessible text will require more sophisticated learning processes on the part of students. Unfortunately, the World Wide Web, although routinely touted as a breakthrough technology for educators, poses special difficulties with respect to the process focus advocated by content literacy experts (see, e.g., Anderson & Joerg, 1996; Edwards & Hardman, 1989; Nielsen, 1989). Fortunately, new dynamic capabilities of on-line reading environments and the tools they support should help software designers and educators to develop learning materials that can help readers avoid some of the problems with web-based content that have been documented,in the literature. The objective of this article is to briefly describe and, in the path-based version, to illustrate by example how new web technologies can be applied to assist readers both in
The purpose of this research was to investigate the role of Internet and computer experience on the selection of vertical or horizontal navigation, the time to complete information-gathering tasks on the Internet, and the frustration level associated with an Internet-search task. The majority of users selected vertical navigation, but there were no statistically significant difference in time to complete tasks associated with either navigational orientation. There was a difference in time associated with experience level: subjects with more Internet and computer experience completed tasks faster. The implications for web designers debating vertical vs. horizontal navigation are discussed.
In this paper we present the design and implementation of Visa: a visualization tool for sequence mining over ARM hypermedia presentations. It can synthesize the user clickstream of an ARM hypermedia environment as to get a global view of the displayed documents as well as the displayed relevant media resources in it. Using Visa, user can get a deep understanding of what is of interest, and thus, how to reach more attendance in hypermedia delivery. We introduce in this paper the requirements in ARM context. Our sequence mining technique is based on distance comparison of relational attributed graphs. This graph represents the user navigational path through documents or through resources that are referenced in these documents
Solving real-world problems is an effective learning activity that promotes meaningful learning in formal educational settings. Problems can be classified as being either well structured or ill structured. Internet information search approaches have an influential role to play in the successful performance of problem solving. A better understanding of how students differentially model information search strategies and movements in tackling well- and ill-structured problems is essential for creating engaging problem-solving environments for students. Static measures, such as the number of accessed nodes or links, or the number of times particular web browser function buttons are clicked, are limited in their ability to analyze attributes of information search patterns. A more dynamic and spatial representation of web movements and navigational patterns can be realized through the use of navigational paths as data. The two path-specific structural metrics that can be used to assess network-based navigational paths in relation to the structuredness of the problem-solving task are compactness and stratum. These metrics are, respectively, the indicators of the connectedness and linearity of network-based structures defining students’ online navigational visitations during the problem-solving sessions. This study explored the relevance and utility of these two metrics in analyzing the navigational movements of learners in seeking out electronic information to accomplish successful problem solving. The outcome findings of this study show that well- and ill-structured problems demand different cognitive and information seeking navigational approaches. The differing values of the two path metrics in analyzing the search movements organized by students in attending to well- and ill-structured problems were a direct result of the contrasting patterns of navigational path movements. The search patterns associated with well-structured problem solving tended to be more linear and less connected, whereas those related to ill-structured problem solving were more distributed and inter-connected.
Hypermedia systems may be considered as networks of interconnected fragments of information. In the literature, several attempts have been made to describe characteristics of nodes, links, and of the overall structure of such systems. However, the utility of these attempts is only limited for educational purposes because most of them are uneconomic in measuring and almost none of them is explicitly concerned with learning. In this article, an analogy between information networks and social networks is established. This analogy is used to illustrate the application of standard network analysis procedures. These procedures are available in widespread PC software (e.g., UCINET) and are used for calculating indices mapping structural aspects of hypermedia systems which are hypothesized to be relevant for learning. It is shown how the graphical representation of a hypothetical hypermedia system can be transformed into a matrix format which functions as the basis for computing standard network indices for nodes...
Spatial navigational techniques have been proposed to assist hypermedia users in discerning the structure of the content, facilitating both the organization and integration of new information. In this study, novice hypermedia users tended to use a fairly passive strategy for selecting their movement through the system, based more on screen position rather than deliberate conceptual choices, whether they had a graphical browser for navigation or just hotwords in the text. Most made a deliberate effort to view all screens but often felt confused. Suggestions for improving navigational aids based on path data analysis, survey data, and user interviews are given.
we use the expression "random access instruction" to refer to a cluster of fundamental issues brought into play by nonlinear learning with random access media / this chapter discusses a unified theoretical approach to those fundamental issues, an approach that provides a foundation of principles to guide random access instruction the research discussed in this chapter illustrates aspects of a general theoretical orientation to knowledge acquisition and application in complex content domains argue for the suitedness of this theory to the special needs of random access instruction although this chapter is primarily concerned with the application of theoretical issues of advanced learning on complex topics to a particular instructional approach, there is also a focus on one illustrative context for the discussion: a computer program in the area of literary comprehension and the understanding of complex patterns of individual behavior the project that is the primary focus of this chapter uses a random access (i.e., interactive programmable) videodisc of the complexly structured classic film "Citizen Kane" / the use of programmable video to construct multiple "texts" is intended to prepare learners (better than instruction with a standard linear schema for the film would) to go beyond the ability to merely reproduce the instruction they received, and instead to be able to independently apply the instructed knowledge to new situations (PsycINFO Database Record (c) 2012 APA, all rights reserved)
This study was designed to reveal search strategies in a loosely structured information environment and to examine cognitive and motivational correlates of particular search strategies. Subjects browsed freely through a hyper-card program containing 105 topics on the Vietnam war. Search strategies were evaluated on a continuum ranging from random search to planned, investigative searches. The findings show search strategies to be significantly related to cognitive complexity, academic ability, and need for cognition. More investigative search strategies were also associated with more complex summaries of the information viewed. The search strategies observed are consistent with other studies of search behavior and learning orientations, and tend to validate the role of ability, cognitive complexity and need for cognition as predisposing variables influencing behavior in loosely structured information settings.
This article examines the effect of learning style on learner performance within two different hypertext structures. Each instructional environment is modeled by a Computer-Based Instruction (CBI) on the Shortest Path Algorithm. Both CBI applications have the same information but different structures for accessing the information. One CBI application with a hierarchical structure provides access to a concept only when all the subordinated concepts have been accessed. The other CBI application with a network structure provides access to any related concept with no limitations. Learning style was categorized using Kolb's Learning Style Inventory (LSI). An experimental study conducted with forty students revealed that both applications were equally effective, and were able to accommodate learners with different learning styles. The study also found that learners using network structures completed the instruction sooner than learners using hierarchical structures, and there was no significant difference in their achievements.
Hypertext documents are unlike traditional text for not only do they represent a computer-based medium, but also readers can process the hypertext in a nonlinear, random access fashion. The ability to navigate through hypertext sometimes presents individuals with difficulty perhaps because these readers do not possess adequate domain knowledge or interest in the topics encountered. This study was designed to inspect the navigational profiles of participants as they process a hypertext document. Using cluster analysis, three performance profiles emerged: 1) knowledge seekers, 2) feature explorers, and, 3) apathetic hypertext users. Analyses demonstrated that domain knowledge seems to differentiate among the cluster groups. Results indicated that students who are interested in computers and hypertext but who do not possess relevant amounts of domain knowledge aligned with the text material experienced difficulty when trying to comprehend hypertext. Implications for learning, assessment, and teaching are discussed.
As part of the general study of user navigation through interactive databases, a set of indices is proposed to characterise users' search sequences. These indices are presented as an attempt to capture some of the psychologically significant aspects of the movement of users within interactive data-bases. An example of the use of the indices is given in reference to an experiment to compare front-ends to a data-base program. Their strengths and weaknesses are discussed in the light of this experiment. Their potential as a basis for the precise definition of search strategies such as browsing and scanning is also considered.
Discusses audit trails (i.e., records of learner responses) in interactive multimedia and hypermedia instruction. Topics addressed include four purposes of audit trails, including formative evaluation in instructional design, basic research, usage studies, and counseling and advising; advantages and limitations of the audit trail approach; and further research needs. (21 references) (LRW)
Examined the performance and navigation strategies of 29 college students engaged in a question–answering task using either a hypertext encyclopedia of Sherlock Holmes facts or its traditional paper form. Results showed that the hypertext users were marginally more accurate in answering questions and excelled at questions where the key information was embedded in a text entry. Book users were marginally faster, overall, but excelled only in answering questions based on graphics. Hypertext users showed a preference for those tools which most closely mimicked use of a conventional book. They used a hierarchical structure to guide their navigation strategy in early trials, but soon learned to navigate in a nonhierarchical way. (PsycINFO Database Record (c) 2012 APA, all rights reserved)