Eye Tracking and Visual Analytics
Eye Tracking and Visual Analytics
Eye Tracking and Visual Analytics
Series Editors:
K.C. Chen Sandeep Shukla
National Taiwan University, Taipei, Taiwan Virginia Tech, USA
and and
University of South Florida, USA Indian Institute of Technology Kanpur, India
The “River Publishers Series in Information Science and Technology” covers research
which ushers the 21st Century into an Internet and multimedia era. Multimedia means the
theory and application of filtering, coding, estimating, analyzing, detecting and recognizing,
synthesizing, classifying, recording, and reproducing signals by digital and/or analog devices
or techniques, while the scope of “signal” includes audio, video, speech, image, musical,
multimedia, data/content, geophysical, sonar/radar, bio/medical, sensation, etc. Networking
suggests transportation of such multimedia contents among nodes in communication and/or
computer networks, to facilitate the ultimate Internet.
Theory, technologies, protocols and standards, applications/services, practice and
implementation of wired/wireless networking are all within the scope of this series. Based
on network and communication science, we further extend the scope for 21st Century life
through the knowledge in robotics, machine learning, embedded systems, cognitive science,
pattern recognition, quantum/biological/molecular computation and information processing,
biology, ecology, social science and economics, user behaviors and interface, and applications
to health and society advance.
Books published in the series include research monographs, edited volumes, handbooks
and textbooks. The books provide professionals, researchers, educators, and advanced students
in the field with an invaluable insight into the latest research and developments.
Topics covered in the series include, but are by no means restricted to the following:
• Communication/Computer Networking Technologies and Applications
• Queuing Theory
• Optimization
• Operation Research
• Stochastic Processes
• Information Theory
• Multimedia/Speech/Video Processing
• Computation and Information Processing
• Machine Intelligence
• Cognitive Science and Brian Science
• Embedded Systems
• Computer Architectures
• Reconfigurable Computing
• Cyber Security
For a list of other books in this series, visit www.riverpublishers.com
Eye Tracking and Visual Analytics
Michael Burch
University of Applied Sciences, Chur, Switzerland
River Publishers
Published, sold and distributed by:
River Publishers
Alsbjergvej 10
9260 Gistrup
Denmark
www.riverpublishers.com
Preface xi
1 Introduction 1
1.1 Tasks, Hypotheses, and Human Observers . . . . . . . . . . 3
1.2 Synergy Effects . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Dynamic Visual Analytics . . . . . . . . . . . . . . . . . . 11
2 Visualization 17
2.1 Motivating Examples . . . . . . . . . . . . . . . . . . . . . 19
2.2 Historical Background . . . . . . . . . . . . . . . . . . . . 27
2.2.1 Early Forms of Visualizations . . . . . . . . . . . . 28
2.2.2 The Age of Cartographic Maps . . . . . . . . . . . . 30
2.2.3 Visualization During Industrialization . . . . . . . . 32
2.2.4 After the Invention of the Computer . . . . . . . . . 34
2.2.5 Visualization Today . . . . . . . . . . . . . . . . . 36
2.3 Data Types and Visual Encodings . . . . . . . . . . . . . . 38
2.3.1 Primitive Data . . . . . . . . . . . . . . . . . . . . 39
2.3.2 Complex Data . . . . . . . . . . . . . . . . . . . . 42
2.3.3 Mixture of Data . . . . . . . . . . . . . . . . . . . . 48
2.3.4 Dynamic Data . . . . . . . . . . . . . . . . . . . . 50
2.3.5 Metadata . . . . . . . . . . . . . . . . . . . . . . . 52
2.4 Interaction Techniques . . . . . . . . . . . . . . . . . . . . 53
2.4.1 Interaction Categories . . . . . . . . . . . . . . . . 54
2.4.2 Physical Devices . . . . . . . . . . . . . . . . . . . 58
2.4.3 Users-in-the-Loop . . . . . . . . . . . . . . . . . . 61
v
vi Contents
3 Visual Analytics 75
3.1 Key Concepts . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.1.1 Origin and First Stages . . . . . . . . . . . . . . . . 78
3.1.2 Data Handling and Management . . . . . . . . . . . 79
3.1.3 System Ingredients Around the Data . . . . . . . . . 86
3.1.4 Involved Research Fields and Future Perspectives . . 88
3.2 Visual Analytics Pipeline . . . . . . . . . . . . . . . . . . . 91
3.2.1 Data Basis and Runtimes . . . . . . . . . . . . . . . 91
3.2.2 Patterns, Correlations, and Rules . . . . . . . . . . . 93
3.2.3 Tasks and Hypotheses . . . . . . . . . . . . . . . . 97
3.2.4 Refinements and Adaptations . . . . . . . . . . . . 102
3.2.5 Insights and Knowledge . . . . . . . . . . . . . . . 104
3.3 Challenges of Algorithmic Concepts . . . . . . . . . . . . . 105
3.3.1 Algorithm Classes . . . . . . . . . . . . . . . . . . 106
3.3.2 Parameter Specifications . . . . . . . . . . . . . . . 110
3.3.3 Algorithmic Runtime Complexities . . . . . . . . . 111
3.3.4 Performance Evaluation . . . . . . . . . . . . . . . 112
3.3.5 Insights into the Running Algorithm . . . . . . . . . 114
3.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.4.1 Dynamic Graphs . . . . . . . . . . . . . . . . . . . 117
3.4.2 Digital and Computational Pathology . . . . . . . . 118
3.4.3 Malware Analysis . . . . . . . . . . . . . . . . . . 119
3.4.4 Video Data Analysis . . . . . . . . . . . . . . . . . 120
3.4.5 Eye Movement Data . . . . . . . . . . . . . . . . . 122
References 273
Index 335
xi
xii Preface
and, in addition, the value of visual analytics for eye tracking. We first
introduce visualization and visual analytics as methodologies to explore
and analyze data with the user-in-the-loop, with and without automatic
analyses and analytical reasoning. This process generates snapshots of
visualizations that support humans due to rapid pattern detection, guiding
further exploration processes like the choice of algorithmic approaches and
applied interactions, and hence helping to build, refine, accept, or reject
hypotheses.
Such visual snapshots – static or dynamic ones – serve as independent
variables in controlled and uncontrolled user evaluations. Typically, those
stimuli are varied, and dependent variables like error rate and completion
times are recorded that are statistically evaluated as a post-process. The
same could be done with eye movement data, although the evaluation is
much more challenging due to the spatio-temporal aspect of the recorded
data and the different stimuli properties. Moreover, additional data sources,
and qualitative feedback, come into play, making such an analysis even
more complicated. However, using visual analytics such heterogeneous data
can be made explorable, in the case where right visualizations, interactions,
and algorithmic approaches are chosen, also allowing human users to
collaboratively and remotely identify insights, sharing them with others, and
combining them into even stronger and more valuable insights.
Visual analytics combined with more advanced data science concepts
like machine learning can be used to analyze recorded eye tracking data,
either offline, after the recording of the data, or online, during the recording,
making it a real-time evaluation process. The insights gained from these
rapid analyses can be applied to the shown stimuli in order to improve
them or adapt to the observers’ requirements and needs. Consequently, visual
analytics plays a crucial role, since it contains many useful methods for
tackling upcoming challenges, although some are very hard and belong to
future work. We conclude this book by several open problems in the field of
eye tracking in general, but also in visual analytics applied to eye tracking
data in particular.
List of Figures
xiii
xiv List of Figures
xxix
List of Abbreviations
2D Two-dimensional
3D Three-dimensional
AI Artificial Intelligence
ANOVA Analysis of variance
AOI Area of interest
AR Augmented reality
BCE Before the common era
CP Comparative
CT Controlled
DBLP The Digital Bibliography & Library Project
DNA Deoxyribonucleic acid
ECG Electrocardiogram
EEG Electroencephalography
EMG Electromyography
ETRA Symposium on Eye Tracking Research & Applications
GGobi Free statistical software
GPS Global positioning system
GUI Graphical user interface
HCI Human-computer interaction
HD High definition
Hz Hertz
IEEE VIS The premier forum for advances in visualization and
visual analytics
JASP Open source statistics program
KDD Knowledge discovery in databases
LB Laboratory
MATLAB Programming and numeric computing platform
MDS Multidimensional scaling
MRI Magnetic resonance imaging
MTurk Mechanical Turk
NASA National Aeronautics and Space Administration
xxxi
xxxii List of Abbreviations
1
2 Introduction
Figure 1.1 Building, rejecting, confirming, and refining of hypotheses plays a key role in
visual analytics.
4 Introduction
(a) (b)
(c) (d)
Figure 1.2 Inspired by “The Unexpected Visitor”: a different kind of visual attention is
reflected in the scanpaths depending on the tasks the spectators have to solve as given in
scenarios (a), (b), (c), and (d) [539].
the appearance of a user interface for example, but for a rapid adaptation,
a real-time analysis of the eye movement data is required. However, the
scanning differences might not only be caused by the tasks at hand but also
by the experience level of a person who was eye tracked. Consequently, by
just algorithmically deciding or predicting, such a task could become error-
prone, hence a combination of visual analytics and a human observer can be
a great way to achieve more reliable results. This again shows the value of
eye tracking for visual analytics [306] but also the value of visual analytics
for eye tracking [14].
Considering the reliability of the results brings the number of eye tracked
people into play. There is no clear definition for that in the corresponding
literature, but since the recorded eye movement data is quite complex,
consisting of several data types, the number of study participants cannot be
large enough. One rule might be the more the better; however, we have to take
6 Introduction
into account the applied algorithms for the data analysis and the fact that the
generated visualizations can suffer from scalability issues, i.e. algorithmic as
well as visual scalability issues in this special case. Scalability means that the
algorithms and visualizations in use can handle larger growing datasets, for
example, growing in logarithmic or linear time with the increase of the data.
Typical scenarios for algorithms are much worse, for example, when they
fall into the class of NP-hard problems [195], or when they have quadratic
or cubic runtime complexity. Heuristics are required in this case, although
the generated results are no longer optimal. But, as a negative consequence
of a poorly performing algorithm, the interactive responsiveness of the visual
analytics system will also suffer. Who wants to wait for a few seconds, or even
minutes, for a clustering algorithm to produce an optimal visual grouping?
For visualization techniques we are typically confronted by the visual clutter
problem [426] if the data to be visualized grows too large. This effect is
regarded as “the state in which excess items or their disorganization leads
to a degradation of performance at some task”.
The number of the involved eye tracked people is definitely a challenge,
either for recruiting them or for analyzing their scanpaths over space and
time. In addition, it is questionable whether the eye movements can be
really used or if privacy issues and ethical reasons make an algorithmic and
visual analysis problematic or even impossible, in particular when sharing
the results with other people. This can even be a major problem if the data
is anonymized due to the fact that at least some of the unknown data might
be recovered by clever algorithms, hence also eye movement data has to be
taken with care. Eye movement data recorded in future user interfaces might
be used to manipulate people’s decisions, i.e. when we know where people
are paying visual attention to. Privacy can even be an issue for the provided
stimuli, for example if other people serve as stimuli to be watched by human
observers to explore where we pay visual attention to, with or without tasks
to be solved. For example, which regions are visually observed for certain
facial expressions might be an interesting study, including infants. Moreover,
dental imperfections might be worth investigating from a visual attention
perspective [275] (see Figure 1.3).
As a negative consequence, privacy issues lead to complications when
a large number of eye tracked people are needed with various additional
measures, for example, to get reliable, statistically significant results to
successfully adapt a scenario or improve a design flaw, even more if this has
to happen in real-time. A large number of scanpaths would be required for
1.2 Synergy Effects 7
Figure 1.3 Human faces can serve as visual stimuli in an eye tracking study, for example to
identify dental imperfections [275]. Image provided by Pawel Kasprowski.
tasks like car driving or shopping, however, smart phones might offer a way
to provide various datasets, just like in a crowdsourcing experiment.
Figure 1.4 Software functions calling each other can change over time and can create a large
and time-varying relational dataset worth investigating by a software engineer to identify bugs
or performance issues in a software system [67].
Figure 1.5 An illustration of the synergy effect. Standard/traditional user studies or eye
tracking studies are conducted while the recorded data is statistically evaluated or explored
by visual analytics concepts.
Figure 1.7 Cognitive science and psychology are also important research fields to improve
the design of eye tracking studies and the interpretation of the recorded data [305].
of the major concepts that are most powerful for detecting valuable claims
about human behavior applied to provided stimuli, static or dynamic ones, in
case they are combined in a meaningful way. This way, tasks solved in an
eye tracking study are not only dependent on the shown stimulus but also on
important aspects like cognitive load [512] in the cognitive processes [268]
or the working memory [386] which is typically not seen in the recorded eye
movement data. The so-called eye–mind hypothesis [269] plays a role in this
context, stating that people visually fixate what they process, having its origin
in reading research [268]. Also machine learning, statistics, and data science
play a crucial role in finding insights, patterns, correlations, and knowledge in
the eye tracking data efficiently, also for classifying scanpaths or predicting
visual behavior based on existing data.
Figure 1.8 Dynamic visual analytics describes the process of evaluating a visual analytics
stimulus by either a post process analysis or a real-time analysis. In a post process analysis,
a second visual analytics system can be used to analyze the eye tracking data; in a real-time
analysis, efficient algorithms must be used.
data for patterns and changing the visual content when required based on the
output of the efficient algorithms. In a second concept, visual analytics can be
used to analyze the data recorded in an eye tracking experiment with another
visual analytics system as a visual stimulus [46]. This concept typically does
not work in real-time due to the fact that the human user is involved in visual
analytics and the eye tracking data has to be analyzed by a combination
of interactive visualizations, algorithms, and the human observers to build
hypotheses about the recognized visual patterns. The visual content of the
static stimulus or a temporal snapshot of the dynamic stimulus is important
to guide the observer, also leading to different applications of interaction
techniques or varying parameter settings for the applied algorithms or to using
completely different classes and variations of supported algorithms.
Exploring the recorded eye tracking data using a second visual analytics
system allows insights into the data, but it also gives us information about
how well the second visual analytics system works as some kind of user
evaluation. This information, on the other hand, can be used to find design
flaws in the second system while the actual task was to adapt or improve
the first visual analytics system, i.e. the original dynamic stimulus in the
eye tracking study. Creating such a dynamic visual analytics system sounds
like an easy task to tackle, however, it involves several challenges to be
solved. For example, a visual analytics system can also be equipped with
1.3 Dynamic Visual Analytics 13
(a) (b)
Figure 1.9 Two types of interactive stimuli: (a) a user interface of a ticket machine and
(b) a more complex user interface of a visual analytics system. Image provided by Bram
Cappers [106].
Figure 1.10 Coordinating multiple views provides several perspectives on the data under
exploration. In this case we see a visual analytics tool for analyzing eye movement data from
video stimuli [308]. Image provided by Kuno Kurzhals.
17
18 Visualization
(a) (b)
Figure 2.1 (a) Part of a visualization for dynamic call graphs [75]. (b) A hierarchy
visualization based on the space-reclaiming icicle plots [509].
a crucial role in the rapid pattern finding process which typically makes
visualization a suitable concept, in many cases performing better than if pure
algorithmic approaches are applied that require an exact definition of the input
parameters. This means that visualization can be helpful in situations in which
the data analysis problem cannot be defined in enough detail. The human user
decides which parts of the data are interesting and require further attention,
given that the chosen visualization is an adequate one.
Typically, the human observer is not left alone with the static diagram,
but interaction techniques [476, 544] awaken the otherwise static visual
representation, providing us with several other perspectives on the data, even
in multiple coordinated views [374], linking various views and giving us
the full potential a visualization can have. All this can only be achieved by
good design of the whole graphical user interface [461] in which the linked
views are laid out as well as the design, layout, and arrangement of the visual
elements in each of the visualization techniques provided in each individual
view. Hence, it is important to have a good understanding of and some
experience in visualization before starting to develop advanced visualization
tools. But it is also crucial to have knowledge about the concepts of computer
and data science, like powerful algorithms that transform or project data into
a pattern-preserving and usable form. The efficiency and effectiveness of the
incorporated algorithms are important to achieve responsive interactions that
allow the views to be adapted and to rapidly inspect the data from different
visual perspectives.
Performance evaluation explores how fast the individual algorithms are
and whether certain algorithmic scalability issues might occur that make
real-time interactions hard to be applied [195]. In this case a better data
preprocessing and data handling stage might be required to first bring the
2.1 Motivating Examples 19
data into a form that allows efficient operations, making the interactive
visualization tool acceptably fast. User evaluation [397], in particular, if
many users are involved like in crowdsourcing experiments [52], with or
without eye tracking, can provide further insights into the usefulness of a
visualization tool, i.e. if there are design flaws that hinder spectators from
quickly getting insights into the data. For example, a chosen visualization
technique might be more difficult to understand than another equivalent one.
This has already been evaluated for simple graphical representations [124].
Moreover, the arrangement or layout of visual components or views in a
graphical user interface might not be perfectly chosen for the tasks at hand.
Even visual variables like color coding or font sizes might cause problems in
the data exploration process, which can be detected by user studies; however,
the parameter space is so huge that not all aspects can be studied in such
experiments, even the simplest ones. Positively, this makes user evaluation
a research topic on its own with all its facets and variations. Eye tracking
brings an even bigger challenge into play since it produces spatio-temporal
data that is hard to analyze for patterns focusing on improving visual designs.
But, on the other hand, it provides a great opportunity to dig deeper into
visual attention behavior and hopefully combining eye tracking successfully
with cognitive psychology [305] one day, to tap the full potential of user
evaluation.
Figure 2.2 Connecting the list of 2D coordinates from Table 2.1 by straight lines in order
transforms the raw data into a visual form, in this case a pentagon shape.
Figure 2.3 The shape of a pentagon as in Figure 2.2 can be drawn with just five points and
five connecting lines, but the shape of a Christmas tree requires many more points. There are
many more complex pattern examples in visualization.
Figure 2.4 Connect-the-dots is a popular game in newspapers and magazines. The human
brain needs lines connecting the dots to successfully interpret the shape.
The created shapes we saw here were just simple data examples consisting
of a handful of points, easily manageable by hand with pencil and paper.
Visualization these days, supported by a computer, has the benefit that it can
visually encode hundreds, thousands, or even millions of data points very
exactly in a short space of time, in the best case reflecting hidden patterns
in the data. In addition, if the visualization of all the data points at once is
not possible due to space limitations or a lot of overdraw and visual clutter,
aggregation or projection methods can be applied to reduce the amount of
data while preserving most of the hidden patterns, but at the cost of a loss of
information.
If we have never seen the pattern of a pentagon or a Christmas tree we
are not able to interpret it, or communicate it to someone else because we
have no common term for it. Visualization is hence based on a common
language of known patterns, just like letters that form words which form
sentences and, finally, an entire story, but the repertoire of visual patterns can
easily be extended (the number of letters is fixed). However, the visual pattern
repertoire must be extended in everybody’s pattern repertoire, otherwise we
cannot easily discuss the findings and communicate the insights. Nearly any
kind of shape can be generated from a list of 2D coordinates and it depends
on the repertoire of known patterns as to which ones can be successfully
interpreted and are useful for a suitable communicative visualization.
There is a larger list of visual variables that can be used to build any kind
of complex visual representations of data. A visual variable is a container for
a certain visual value that can change, but that still has the same type as all
the values fitting in the same visual variable. A visual variable describes to
which visual elements of the same type a data value is mapped. An example
of a visual variable is the color hue. The hue can be changed but it is still a
2.1 Motivating Examples 23
Figure 2.5 Visual variables are fundamental ways to distinguish visualizations. Jacques
Bertin [37] described seven such variables and denoted them as retinal variables.
hue. Examples for visual variables are illustrated in Figure 2.5 and are color,
size, position, shape, value, orientation, arrangement, texture, and several
more [39, 344]. For color we can have a finer categorization into hue, value, or
saturation for example. There are also some more modern ones like crispness,
resolution, or transparency [342].
Before starting to design and implement complex visualizations, it is
important to understand the fundamentals or the basic rules in visualization
first. Two pioneers in this domain were William S. Cleveland and Robert
McGill who tried to figure out which kinds of visuals lead to good
performances in terms of error rates. In their experiments [124], they showed
that some visual variables are better than others for a given task in terms of
performance, like error rates. However, they were not applying eye tracking
to record the visual attention paid to their stimuli. The spatio-temporal eye
movement data would have given them many more insights, in which case
they would have been able to analyze it. Eye tracking devices and also
visual analytics were either not that well researched or even did not exist
as they are known today in their advanced forms. The steady progress in
hardware technologies and software engineering has led to better, faster, and
more efficient solutions. However, what they found out with traditional user
experiments was the fact that the visual variable position in a common scale
leads to better user performance than using the angle for the task of judging
the size of a data point for example.
This effect appears in the very common and simple visualization
techniques like bar charts and pie charts, which were used as polar area
diagrams by Florence Nightingale during the Crimean War (1853–1856) to
show deaths in British military hospitals [130], although it is said that they
were invented much earlier by William Playfair [188] (see Figure 2.6). Pie
charts are very common representations in newspapers or magazines for the
24 Visualization
(a) (b)
Figure 2.6 The task of judging and comparing sizes appears to be easier and more reliable
with fewer errors in bar charts (b) than in pie charts (a). Visual variables are the cause of this
effect, which are positions in a common scale in bar charts and angles in pie charts.
all of the value lists are the same, or at least similar. The benefit of the
derived values is that we get rid of the large number of values by aggregating
them into one number; however, aggregation is always prone to data loss.
Statistics is a powerful concept, but only applying statistical approaches to
detect insights into a dataset is a process only showing half of the truth and
can lead to various misinterpretations of the underlying data.
Representing the values from Table 2.2 in 2D scatterplots and inspecting
their distributions, i.e. visual patterns, reveals something totally different than
reflected by the pure statistics. The visual depiction of this simple bivariate
dataset provides more insights than the statistical summaries into individual
aggregated values (see Figure 2.7). This does not mean that statistics is
useless or error-prone, but a look at a corresponding visualization can help
to see the data under a different light, maybe leading to more detail and more
fine-granular hypotheses. This effect was studied by Anscombe [15], and is
known as the Anscombe quartet.
Most of the data situations do not allow such a simple depiction
as a scatterplot, but rely on more complex visual scenarios, inspired by
26 Visualization
the famous saying “a picture is worth a thousand words” [399], which first
appeared in a 1911 newspaper article [57].
However, although the creation of visualizations is definitely possible
without devices such as the computer [37, 39], it would probably take much
more time to generate them, in particular if the underlying dataset consists of
a variety of data entities. Moreover, the visual result will be static, meaning
it cannot be modified in a fraction of a second, keeping pace with the
perceptual abilities of the human brain to rapidly detect patterns. In particular,
many of those visual patterns, watched from various perspectives make a
visualization tool a powerful concept to keep up with the vast amounts of data
generated in these days. Without the use of the computer combined with the
experience and knowledge of human users, much of the data would hide most
of the patterns and hence the informational insights that are contained in it.
Consequently, the visualizations designed and sketched before the invention
of the computer could not tap the full potential for data visualization of the
big data era that we find today, consisting of various data sources, being static
or changing over time.
In this section we will focus on the history of visualization, giving a brief
overview of how the field has developed over time to get a rough impression
of why the application of eye tracking to visual concepts is very important but
also challenging due to the fact that many more visual variables are combined,
while interactions allow us to quickly switch between them based on users’
demands.
Figure 2.8 The outline of an animal found in the Cave of Altamira near Santander in Spain,
also known as the Sistine Chapel of the cave painting.
In these prehistoric times, the human eye also played a crucial role, for
example, when rapidly detecting a predator hunting for food. The strength of
human sight was one key to survival in the wild, for example during hunting
it was important to rapidly distinguish the berries in a bush from the eyes of
a wild animal like a tiger with food in mind. Touch, hearing, smell, and taste
were also important in these times but no sense has a higher bandwidth than
sight transmitting information faster than any other sense to the brain to react
on situations that might possibly have had a bad impact on a human life.
Apart from inspecting natural scenes, the human eye was also important
for interpreting behavior among humans, by seeing where a gaze was directed
one could guess the mood of another human being. In this context, the so-
called cooperative eye hypothesis [290] was researched as a means to describe
visible characteristics to facilitate humans when following another human’s
gazes while they communicate simultaneously, or while they collaboratively
solve certain tasks. The research in this field was guided by scientists
from the Max Planck Institute for Evolutionary Anthropology in Germany
investigating effects of head and eye movements on gaze direction changes,
comparing humans and great apes.
In particular, for the cave paintings, some theories have been developed
over the years. One of them regards the paintings as a means of
30 Visualization
(a) (b)
Figure 2.9 Maps have been used in a variety of forms, including various visual variables:
(a) a geographic map annotated with a grid-based overlay to faster detect the label information
and the location of a place [371]. (b) Data from other application domains with a more abstract
character have been visually encoded into maps, like trade relations [243]. Figure provided by
Stephen Kobourov.
some kind of user evaluation is needed to investigate whether people can still
interpret the data visually to solve the tasks at hand. Eye tracking could be
useful here, since it brings into play spatial visual attention behavior that can
be visually encoded directly on the map to see where the design flaws or
problems might occur during paying visual attention [371].
Figure 2.10 Today’s pie charts are based on the ideas originally developed, for example, by
Florence Nightingale.
the application field or topic that the infographic is about, sometimes denoted
by the term reference graphic. Although it is said that they originated many
centuries ago, they had their high tide in the years of industrialization due
to the fact that the data was quite small and no computer was available to
quickly produce or print them. However, they are still created in the 21st
century because they typically attract the attention of the human observers
and are quickly, easily, and effortlessly understood.
There was also no kind of advanced user evaluation giving feedback
about the usefulness of the diagrams, about their interpretability as well
as understandability and effectiveness, for example taking into account the
encoded visual variables or perceptual issues. However, it seems that user
evaluations, even for the simplest infographics, are conducted today as
progress in technology and a growing visualization community, in particular
eye tracking studies [346], become more and more of interest, since they can
uncover design flaws over space and time, not only aggregated to measures
like response times and error rates as in traditional user experiments.
There are various hand-made examples from the early days of
infographics. For reasons of completeness and for illustrative purposes the
most important ones will be described here. “The best statistical graphic ever
drawn” was a famous quote by Edward Tufte that illustrated the expressive
power in the so-called Minard map [503]. The graphic designed by Charles
Joseph Minard (1781–1870) shows Napoleon’s army marching to Moscow
34 Visualization
Figure 2.11 An example of a graphical user interface for visually exploring eye movement
data [82]. Figure provided by Neil Timmermans.
slightly adapt it to the needs given by the newly computed structure of the
data and the users’ tasks at hand. The amazing power of the computer brings
a new topic into play but also even more new challenges: real-time data
visualization [533]. Not only is real-time data a challenge for visualizers but
also data that is changing over time, allowing to play back or replay several
times, i.e. after the dynamic data has been recorded and preprocessed.
We discussed time-dependent data in infographics in Section 2.2.3, for
example the march to Moscow and the retreat, which has an implicitly
inscribed timeline, but shows the same underlying data as an animated
diagram with the aforementioned features. This was impossible in the years
before the invention of the computer, at least not with an equal outcome.
Moreover, several visualizations could be shown next to each other and
linked, a concept that is known as multiple coordinated views [424]. The
graphical user interfaces (GUIs) in which typical visualization tools are
integrated consist of a variety of additional functionalities like buttons,
sliders, menus, and many more, all targeting the common goal of supporting
users to find insights into a dataset (see for example Figure 2.11 for an eye
movement data visualization user interface). In general, the advent of the
computer meant a huge step in the field of visualization, and the field is
still progressing and outputting lots of research ideas, also based on new
technologies.
The very first definition for the term visualization was given in an
NSF panel in 1987 by McCormick, DeFanti, and Brown [146, 349]
stating: “Visualization is a method of computing. It transforms the symbolic
into the geometric, enabling researchers to observe their simulations and
36 Visualization
Figure 2.12 A hierarchy visualization depicted on a powerwall display. The system allows
collaborative interactions for several users equipped with tablets, or it serves as an overview
of a large dataset [93, 441]. Pictures taken and provided by Christoph Müller.
In this book we describe the most frequently occurring data types that
have been visualized and that have been investigated in user evaluations, in
particular with eye tracking technologies. We start with primitive data types,
proceed with complex data types, and finally explain combinations of data
types as well as time-dependent data, also taking a look at metadata because
it is important for a self-explanatory diagram. In the following we provide
visual examples for representing data categorized by the type to which they
belong. We describe the included visual variables in visual depictions and
discuss the pros and cons whenever we assume that it is required. In some
scenarios we also explain for which tasks a visualization of data of a certain
type is particularly beneficial. The visual encoding and the tasks play a crucial
role in user studies since they play the role of the independent variables for
which dependent variables are recorded, like error rates, response times, or
spatio-temporal eye movement data, in case visual attention during the task
solving process is of interest.
Figure 2.14 Jock Mackinlay gave an ordered list of the visual variables for each of the three
primitive data types. He described the effectiveness of such a perceptual task in decreasing
order [344].
42 Visualization
gets so large that node-link diagrams produce visual clutter [426] and
are no longer readable, even after applying an advanced layout algorithm
focusing on aesthetic graph drawing criteria [406, 409] or changing the
edge representation style [237], and even bundling them together [236,
239]. Hence, adjacency matrices [29] were invented (see Figure 2.15(b))
to get a more visually scalable, clutter-free representation of relational
data, evaluated for typical graph interpretation tasks [200]; however, the
rearrangement of the matrix rows and columns [34] has a large impact
on the identification of patterns, typically vertex groups and clusters. A
third rarely used visualization technique is an adjacency list [229] that
has some benefits but it is quite hard to read paths and identify clusters
(see Figure 2.15(c)). If a graph is locally dense and globally sparse,
hybrid methods are used such as the NodeTrix representation [227].
• Hierarchies are another relational data type with the differences that they
can be drawn as a node-link diagram in 2D in a planar way without
link crossings, that the corresponding graph has no cycles, and the fact
that they have a designated root node. Hierarchical data is built on
parent–child relationships defining the level a vertex is located and also
aspects like its depth and branching factor. Visualizations for this type
of data exist in a variety of forms [445, 446]. For this type of relations
there are also various attributes to be attached to vertices and edges,
like the sizes of files in a hierarchically organized file system or the
evolutionary distances of related species living on earth in a so-called
NCBI taxonomy [509]. For visualizing hierarchical data there exist at
least four major types of visual metaphors (see Figure 2.16) that come
in the form of explicit links [78, 419], nesting [266, 458], stacking [295,
509], and indentation [91, 95]. Each of these types contains various
variants exploiting different combinations of visual variables, even
hybrids of these types and variants exist. User studies, in particular
with eye tracking, can find design flaws and perceptual problems for
typical hierarchy-related tasks like finding the least common ancestor
for example [72, 78].
• Data coming in the form of a table consisting of rows and columns is
said to be multivariate or multi-dimensional data. The rows are called the
observations or cases while the columns are the variables or attributes.
One major task for this kind of data is the identification of correlations,
positive or negative ones, or more complex ones as well as outliers and
anomalies.
44 Visualization
(a) (b)
(c) (d)
Figure 2.16 Four major visual metaphors for hierarchical data exist: (a) explicit links; (b)
nesting; (c) stacking; (d) indentation.
There exist three major visual metaphors for this kind of data which
come in the form of glyph-based representations [325, 428], scatterplot
matrices [172], and parallel coordinates plots [224, 255] that have been
evaluated in several ways in the past [265]. Figure 2.17 illustrates
an example for a multivariate dataset with the same observations
and attributes. We see that the glyph-based representations rely on a
combination of visual variables, one for each attribute, combined in
a single visual entity. The scatterplot matrices consist of a quadratic
scheme of individual scatterplots, each making use of the visual variable
position and maybe color or shape if the points carry categorical
information as well. The parallel coordinates are based on the visual
variables position in common scales and connectedness by straight links,
forming polylines.
• Trajectories are generated by moving objects, animals, or people,
leading to some kind of spatio-temporal data since they change their
locations over space and time. Eye movement data [161, 235] also falls
into this category. If we only take into account a pure scanpath, i.e. the
movements the eye makes, we receive the simplest scenario of such an
eye movement trajectory. In general, eye movement data is much more
complex with extra attributes attached, for example fixation duration,
2.3 Data Types and Visual Encodings 45
(a) (b)
Figure 2.18 Different kinds of movement data can be measured and visualized: (a)
trajectories from bird movement [369]; (b) scanpaths from an eye movement study
investigating the readability of public transport maps [372].
(a) (b)
Figure 2.19 There are various scenarios in which textual information is important: (a) label
information on a public transport map [372] (Figure provided by Robin Woods, Communicarta
Ltd); (b) an aggregated view on the occurrence frequencies of words in the DBLP, summarized
as a prefix tag cloud [86].
Figure 2.20 A set visualization based on the “bubble sets” approach [127]. Image provided
by Christopher Collins.
filter data, navigate in it, or collapse and expand the data to allow more
space for the remaining visible elements.
Figure 2.20 shows an example of a complex visualization based on the
so-called “bubble sets” approach [127]. Numerous visual variables are
used to create fancy but also complex diagrams that require some time
to learn and to interpret the data that is visualized.
• In the field of scientific visualization in particular, we have to deal
with complex data such as 2D/3D scalar, vector, or tensor fields as
well as volumetric data, just to mention a few data type examples
that can easily get more complex by adding more data attributes, and
also time dependency. Moreover, compared to information visualization
the datasets under examination are typically much larger and generate
continuous data values instead of discrete ones, either for the static
spatial data or in the dynamic time-varying case [281].
48 Visualization
Figure 2.22 A part of the Eclipse software system and its hierarchical organization depicted
as a node-link diagram with aligned orthogonal links to visually represent a list of quantitative
values for certain derived attributes [62].
the visual variables for the final visual design, i.e. that decides about the
most effective data-to-visualization mapping. However, in some scenarios
this order is not that clear right from the beginning, offering the users the
choice to reconfigure the visualization in a way that adapts the roles and
orders of the variables in use.
A popular example is the field of software engineering that produces a
wealth of data based on the source code, the involved developers, comments,
bug reports, check-in information, call dependencies resulting in a call
graph, code–developer relations, the hierarchical organization of the software
system, and even more, for example the time-varying aspect bringing the
field of software evolution into play as well as software visualization [149].
In Figure 2.22 we see an example from a software system for which the
hierarchical organization stands in focus, i.e. plays the primary role. The
secondary role is given to a list of quantitative attributes which generate
some kind of multivariate dataset, with the files being the observations. For
the visual depiction we chose a node-link diagram in a top-to-bottom layout
indicating the parent–child relationships while the attributes are reflected
in the color coding represented in an aligned fashion to allow comparison
tasks on different levels of hierarchical granularity, looking similar to bar
codes [62].
50 Visualization
(a) (b)
Figure 2.23 (a) A time-varying graph dataset consisting of flight connections in the US
from the year 2001 shown as a heat triangle [242]. (b) A Themeriver [218] representation for
showing the evolving number of developers during software development [89].
serial visual presentation (RSVP) often used for text reading tasks [113] or
inspecting large image collections [477] in a rapid way. The concept supports
tasks like browsing a large static dataset or a time-varying dataset consisting
of several instances of data values visualized in some way. Moreover,
weighted browsing tasks describe the way in which we search for a certain
element for what we have an approximate visual pattern in mind. No matter
which concept we chose for displaying dynamic data, it is definitely more
challenging to effectively visualize, and further, an eye tracking study has to
take an additional parameter into account which comes in the form of the
sequential behavior [309]. A video might even fall into this category since it
consists of a sequence of static image frames that are watched one after the
other in a rapid serial visual presentation; however, eye tracking experiments
have already been conducted for this kind of scenario [310].
Examples for dynamic data visualizations exist a lot due to the prevalence
of the temporal aspect in nearly any kind of application domain that stores,
measures, or simulates data at different instances or time steps. Figure 2.23(a)
illustrates an example from dynamic graph visualization showing the flight
behavior in the US in the year 2001 depicted as a heat triangle [242].
One can clearly detect a visual anomaly pattern which was caused by the
terror attacks on 9/11. Another popular example (b) stems from the field of
topic exploration which results in a diagram denoted by Themeriver [218]
showing a list of time-dependent quantities (here a developer river [89]), but
52 Visualization
negatively also bringing issues for value comparison tasks into play [124].
Moreover, a simpler diagram showing quantities as a time-series plot is
known as the electrocardiogram (ECG) [144] with which every medical
doctor is familiar (see Figure 2.24). All of the presented diagrams use the
time-to-space mapping concept, i.e. displaying the individual time steps next
to each other to support comparison tasks.
2.3.5 Metadata
Every dataset requires some kind of additional descriptions that explain
and give more details about the data, extra information that is needed to
understand the context of the data. The data about data is important for a
visualization to pick the right visual metaphor, for example the correct scale
given by the units of the measurements. For the user the metadata is crucial
information to interpret the visual patterns in the right context and scale [317].
This data defines how and when data is measured, where it is stored, who is
the owner, in which environment it was measured, which device was used,
and so on. For example, when taking a picture with a camera, the picture is
the data of interest while the date and time when it was taken, the label of
the picture, the resolution, or further extra information, describe the metadata
about it. Moreover, in the field of eye tracking, metadata might describe the
quality of the recordings, giving a hint about how trustworthy and reliable the
data is. This quality can even be measured for every data entity, for example at
different time instances or for every individual participant in an eye tracking
study [444].
For a visualization it is crucial to add the metadata to give the users the
opportunity to interpret what is actually depicted. Without the metadata it
might get misinterpreted or the metadata information must be derived from
another context, making the interpretation process quite time-consuming.
Figure 2.25 illustrates an example of a scatterplot with additional information
about the attributes shown at the axes as well as the color coding of the points.
Without them the visual patterns might be observable but they cannot be
put into the context of an already known scenario with which it should be
compared.
The intention of Section 2.3 is not to introduce and discuss all data types;
that would generate an endless list of examples. It is more given in a way to
provide an overview about the most common data types and possible visual
depictions of them as well as some additional remarks about drawbacks and
benefits of one technique compared to another one. Later, these techniques
2.4 Interaction Techniques 53
Figure 2.25 A scatterplot with extra descriptions for the axes and the color coding, serving
as metadata.
will be examined for user evaluation, in particular if eye tracking was used to
investigate their usefulness, readability, and interpretability by human users
based on visual attention paid over space and time, hence explaining the value
of eye tracking for these techniques.
Figure 2.26 A visualization pipeline illustrates how raw data is transformed in a stepwise
manner into a graphical output while the users can adapt and modify the steps and states [108].
Figure 2.27 An interaction history can be modeled as a network of states; in the case of a
visualization tool it consists of snapshots, each illustrating a certain parameter setting [68].
Node-link diagrams can depict the weighted state transitions and are interactive themselves.
Here we see a network with additional thumbnails indicating the current view in the
visualization tool. It may be noted that in this scenario some of the interaction steps cannot
be undone, indicated by the directions of the links while some others are undirected. The
thickness of the links might encode a transition probability for example.
graph consisting of states and transitions, meaning also for this kind of dataset
we need an advanced interactive visualization technique, otherwise it would
not be useful at all.
A further interaction technique can be the editing of a textual description,
like changing a label or adding an annotation, although annotations could
be classified as connect interactions since they link a textual description to
a visual element. Such a modification, update, or even correction of text
could even be applied directly to visual elements, for example by exchanging
a visual variable or removing a visual pattern that is superfluous which
would again fall into the category of encode. Moreover, more visualization-
independent interactions might be imaginable like adding extra data sources
or requesting a data update if the visualization tool is not using the latest
version. However, such an update could demand a longer preprocessing time
to bring the new or missing data into the same format as the already loaded
one. Even more visualization-independent interactions might be opening
and closing menus, menu items, or whole views containing a visualization.
Also interactions related to the graphical user interface in which all the
visualizations and their functionality live can build their own category; for
example, maximizing or rearranging views, adding more views, highlighting
or decorating a view with additional information, and many more.
The way in which we interact in a visualization also depends on
further factors, for example, based on the visual variables directly. A 2D
interaction is typically different from a 3D interaction in which occlusion
and perspective distortion effects have to be taken into account. This could,
in particular, be interesting and challenging for VR/AR environments; for
example, in immersive analytics, a field that is more related to visual
analytics since it supports complex exploration processes including many
technological concepts. Furthermore, interacting with a static visualization
requires different interaction techniques than when interacting with animated,
dynamic visualizations. Those need to be stopped, replayed, slowed down, or
sped up on user demand.
In these modern days there exist various input devices while the number of
output devices is not that large in comparison. Evaluating the impact of input
devices on the usability of a visualization tool by eye tracking is similarly
challenging as evaluating the impact of the output devices. However, for
the output devices we obtain the primary stimulus, i.e. most of our visual
attention is paid on the output (the visual representation) rather than on
the input, hence most of the recorded eye movement data contains output
device information. Typically, the information from the input device has to be
collected by additional sensors, apart from gaze-assisted interaction in which
the input data comes from the eye.
Most of the input devices focus on hand control, for example mouse,
keyboard, touch, pen, or joystick, to mention a few. For touch devices, the
input and output typically happens on the same display, for example the
computer monitor or a mobile phone to provide a faster and more natural
feeling during interacting. Touch supports direct connection of the user’s
finger with a graphical element while a mouse does not and requires some
synchronous action with the eye and the hand to efficiently work. Hence,
mouse interaction needs some training compared to touch interaction. Touch
is, these days, integrated in most of the public user interfaces, for example, in
ticket machines, information monitors, interactive public transport maps [96],
and so on. However, for a visualization technique it might lead to occlusion
problems due to the size of a human finger compared to a mouse cursor, for
example. This could, in particular, be a problem for eye movement studies
because the visual attention behind the finger is hard to track reliably.
Apart from hand-driven interactions, further input modalities have been
researched and developed including gestures, voice, or gaze [381]; even a
combination thereof can be efficient if it is designed in a user-friendly way.
For example, for gaze interaction, which is useful for disabled people who
cannot just use standard interactions; the popular Midas touch problem [260,
516] typically makes a pure gaze-assisted interaction challenging, hence
some approaches combine gaze and voice interaction (see Figure 2.28).
For game playing such a setup was evaluated, comparing gaze and voice
interaction with mouse and keyboard [378]. For gaze-assisted interaction,
an eye movement data evaluation can be started right away because the
eye movement data is already recorded which is typically not the case
if just a standard non-gaze-based user interaction is supported. For even
more complicated settings such as 3D interactions [191], for example on
mobile devices [104] as well as walking in 3D immersive virtual reality
environments [145] or tangible visualization [287], standard interaction
60 Visualization
Figure 2.28 Gaze-controlled buttons are used in a game environment to interact with a game
character. After a user evaluation this was replaced by a simpler scenario due to unintended
rotation issues [378]. Image provided by Veronica Sundstedt and Jonathan O’Donovan.
2.4.3 Users-in-the-Loop
Human users are responsible for initiating interaction techniques. They awake
a visualization to life, guided by tasks at hand that require some kind of
solution, in most cases raising new questions falling in a certain task category.
The human users cannot only intervene in the final visual output but in typical
visualization tools, they have the option to interact at any stage illustrated
by the visualization pipeline (see Figure 2.26), making them users-in-the-
loop. This pipeline can even be run through in an iterative manner, adjusting
and adapting parameters, views, and perspectives at any stage, until the final
desired configuration is hopefully reached, generating an “aha” effect, i.e.
either providing a solution to a task or at least giving a hint where to search
further. Interaction can be subdivided into subinteraction sequences in which
all interactions should occur quickly, typically in a fraction of a second
from starting the interaction to the final output, i.e. until the interaction has
ended, before a new one from the sequence starts. The human user decides
the order of subinteractions and how this sequence is created, changing
the parameter setting in a visualization tool step-by-step. Some interactions
from the sequence might be executed several times in a row, some might be
reversed or be undone. There is some debate about an acceptable time for an
interaction technique to be considered interactive. Some results indicate that
20 ms [337] might be an acceptable response time, but that surely depends
on the users’ knowledge and experience levels as well as the application.
However, no matter how long it takes for a visualization tool to react on user
input, it is important that a user feels comfortable, engaged, and entertained,
aspects that ease the burden of using a visualization tool with the goal of
deriving knowledge and insights for the tasks at hand. Such properties are
hard to measure with an eye tracking device but additional data sources might
give a hint about them as well as cognitive and psychological issues [305]
worth integrating into a data analytics process to identify design flaws in a
visualization tool.
There is even a difference if an individual user interacts with a
visualization or if the tool supports collaborative interaction [60], i.e. letting
several users work with the tool, either at the same place in front of
the same output device or remotely if they are physically separated, even
one after the other due to different time zones or maybe in temporally
overlapping processes. However, such a scenario typically requires a web-
based visualization tool [82] that offers easy possibilities to store and organize
the results found by many users, all, just a few, or even an individual one.
62 Visualization
values indicating minimum and maximum or they show the individual colors
used for a certain category like, for example, public transport lines.
Axes scales are important, in particular, if an axis is composed of several
scales; for example, if there is a large range between the minimum and
maximum values. In addition, axes descriptions and corresponding units are
required as well as guiding lines that perceptually help to quickly read the
represented value. In a scenario in which the vertical axis is used to indicate
quantities, but those differ in size a lot, a logarithmic or a scale-stack is of
special interest [230], depending on the user tasks. As a negative issue, these
types of stacked scales have to be learned; however, the first attempt to make
them interpretable is by attaching an intuitive scale description.
It is also important that textual information is added in enough detail, for
example labels that give extra hints about scale values. Adding too many is
as bad as adding too little; a good balance is required and texts should be
readable in an acceptable font size and font style. Moreover, a left-to-right
or a slightly varying reading direction should be used for users from Western
civilized countries, for others the reading direction has to be adaptable in
the design. The users’ visual acuity plays a crucial role for text reading
tasks [470], while text reading tasks have been evaluated by eye tracking
a lot in the past [41]. Any additional information is useful, but the diagram
should not be too crowded to avoid an information overflow. However, if extra
textual or visual information is presented, overlaps and occlusion should be
avoided as well as distortions. The choice of the right color scale depends
on the data to be displayed as well as effects related to the user, for example
color blindness [348, 521] or color deficiencies [380].
There could be two extreme situations for data visualization that have to
be treated with care, and this depends on the user tasks. The first issue comes
from the fact that data elements might be missing or erroneous right from
the beginning. Those data elements might be ignored in the visualization or
they might be indicated very clearly to alert the user to those issues. If they
have to be shown in some way, they should be color coded in a gray scale,
indicating that there are missing or wrong values, but they are depicted in
a different way to contextual information. The same holds if those values
are aggregated with other surrounding values or interpolated, and all of these
hints attempt to avoid misinterpretation of the data. A second issue comes
from the fact that some values are considered outliers or anomalies. These
should play a special role in the visualization and should be highlighted in a
way that makes them pre-attentively detectable [219, 500]. This means they
2.5 Design Principles 65
pop out from the display, with just one glance at the monitor, without paying
a lot of attention to them.
No matter which kind of visual enhancements are chosen, the readability,
intuitiveness, and understandability play major roles when designing
powerful visualizations, and aesthetics [24] plays a nearly equally important
role, but should still be considered a minor second option to further improve
a diagram after it fulfills the task solution and explorative functionalities that
a visualization tool should have.
Figure 2.29 The user has the option to see one, two, or four views for hierarchy visualization
techniques. Moreover, the views are exchangeable and support their own parameter settings
while they are interactively linked [99].
66 Visualization
views is adequate, starting with the most important view as the root node and
following a hierarchical exploration task order. In any case the views should
be clearly distinguishable, separable, and a good layering should be used.
The visual information seeking mantra [459] is a good concept to follow
in a visual design. For a data visualization it is important to provide an
overview as a starting point for further exploration processes. After a user
sees the whole dataset or, due to scalability and display limitation issues,
at least a large portion of it [503], further interactions should be supported
like zooming and filtering and finally, details-on-demand. Another option
for reducing the amount of data to be displayed is by using small graphical
elements like pixel-based representations whenever that is possible. For
multivariate data, it could be projected first to a lower dimension and then
visualized, supported by various projection algorithms [176]. In such a
projection process it must be guaranteed that formerly similar data points
in high-dimensional space are still similar in the low-dimensional one to
preserve the patterns and hence, the interpretation of the data in a reliable
way.
No matter how the data is visually depicted, the visualization should
at least try to tell a story [324]. A diagram should be as self-explanatory
as possible [362]; if a lot of textual descriptions are given a good way
to bring them to life by visuals is by using sparklines [429], which are
tiny graphics that are embedded in a text, being so small that even visual
enhancements or decorations are left out. In these modern days, a special
focus could be on a web-based visualization, making it accessible by using a
smartphone, reaching thousands of users; however, the display space is really
small and hence the visualization must be organized in a way that it makes
the most of the small display area. The opposite effect occurs if we need high
resolution, for example showing tiny visual features that are crucial in the
visual depiction to explore the underlying dataset, but for such a scenario the
number of users is typically small, i.e. it is created for domain experts. In
any case the design should be balanced, using symmetrical layouts, to make
it aesthetically appealing, maybe by making use of radial forms instead of
Cartesian ones [24].
usability would be limited in a way that the users cannot really adapt the
presented diagram to their needs by, for example, changing parameters.
No matter which kind of visualization is created, the experience levels and
further properties of the users have to be taken into account to not design
the visualization for the wrong user group, making it useless. Moreover,
interaction is one way to avoid boring visualizations, but even further, visual
variables and in particular color [440] should be chosen in a way to create
enjoyable diagrams, a fact that is also related to aesthetics.
For a static visualization there are already issues that can decrease the
value enormously. Those come in the form of visual clutter which is “a
state in which excess items or their disorganization leads to a degradation of
performance at some task” [426]. This could, for example, happen if line-
based diagrams are used and if there are too many crossing lines like in
a time-series visualization based on line plots (see Figure 2.30(a)). Node-
link diagrams for graph data mostly suffer from this issue. For this reason,
graph drawings have to consider aesthetic rules like reducing the number of
link crossings, avoiding node–link and node–node overlaps, or reducing link
lengths, to just mention a few from a longer list [406, 409]. A data variable
should in the best case only be mapped to one visual variable, otherwise the
additional decoration of a diagram will no longer be a visual enhancement,
but could generate the opposite effect, leading to misinterpretations. Less
is oftentimes better in visualization, leading to some kind of simplistic or
minimalistic diagrams [91]. This is related to the data-to-ink ratio [254, 503]
although this just describes that as little ink as possible should be used for
drawing a diagram, which also includes that second, third, or more visual
variables should not be used for encoding the same data variable, which is
denoted by the term chart junk [26, 503] (see Figure 2.30(b)). Finally, the
lie factor [503] describes the ratio of the extent of the effect shown in a
68 Visualization
visualization and the extent of the effect existing in the data. This proportional
issue typically occurs for displaying comparable values. The chosen visual
variable for the numeric variable should treat each value in the same way, i.e.
a value that is twice as large in the data should also be shown with a twice as
large effect in the visual variable (see Figure 2.30(c) for a strong lie factor).
For dynamic data [4] it is important that two representation choices for
the time dimension can be considered, which are denoted by either time-
to-time mapping or time-to-space mapping [75]. The time-to-time mapping
describes the concept of displaying each time instance in the data to physical
time, for example in an animation, typically called smooth animation in
visualization. Time-to-space mappings try to encode as many time instances
as possible to the display space. While time-to-time mappings generate
cognitive and change blindness [376] problems when comparing time steps
due to limitations of the short term memory [410], time-to-space mappings
typically lead to visual scalability issues due to display space limitations [4].
Another negative issue is that for comparison tasks the individual visual
elements have to be mapped to each other in each time instance to reliably
do the comparison. Moreover, both mappings have to rely on dynamic
stability [151] to allow the preservation of the users’ mental maps [18] which
is even more important in animated diagrams. In non-animated diagrams
we do not speak of dynamic stability but more of a temporal alignment,
meaning that the time axis should be the same for all varying variable values
which leads to better performances for comparison tasks. Both concepts have
their benefits and drawbacks [505] while rapid serial visual presentation
(RSVP) [475] is some kind of hybrid concept that includes ideas from both
concepts.
since they explain which patterns are easily perceivable and which ones could
cause confusions and misinterpretations. This is in particular useful if we
take into account several visual variables of which a diagram is composed.
This does not only hold for static diagrams but also for dynamic, animated
ones that model change by movement, typically causing issues when keeping
track of visual elements, either individual ones or whole groups of them,
merging with each other and splitting off again after some time. The most
important principles are summarized by emergence, reification, multistability,
invariance, and grouping.
The principle of emergence is probably one of the most relevant ones
in visualization since it states that visual patterns might emerge from a
visual depiction, for example, separating them from noise or from chaotic
patterns that do not carry any meaning (see Figure 2.31(a)). Without this
principle the data-to-visualization mapping is useless because we might not
be able to detect visual patterns that can be remapped to data patterns,
hence the visual exploration process would be impossible in this case.
Multistability is important to let the users see several visual patterns from
the same arrangement of visual elements, meaning several perspectives on
the visual depiction of data are possible (see Figure 2.31(b)). The principle
of invariance allows the detection of deformed visual patterns, for example if
they are rotated, stretched, or scaled in any direction (see Figure 2.31(c)). The
perceptual abilities are good enough to recognize similarities which is one
of the strengths of visualizations since a pure algorithmic approach cannot
be used due to the fact that we do not know how to describe the similarity
between two or more visual patterns. Reification is a principle in visualization
that describes how patterns can be completed virtually although they are not
completely visible on screen (see Figure 2.31(d)). For example, a link in a
graph visualization might be dashed or incomplete, or even overplotted by a
larger area.
(a) (b)
(c) (d)
Figure 2.32 There are several ways of grouping visual elements described in the Gestalt
principles. (a) Proximity. (b) Similarity. (c) Closure. (d) Symmetry. Further ones are given by
the law of common fate, continuity, or good form.
(a) (b)
(c) (d)
Figure 2.33 Visual illusions can happen in a variety of forms including visual variables
and the environments in which they are used: (a) Ebbinghaus illusion related to size effect,
caused by the environment and surroundings. (b) Cafe wall illusion related to distance, caused
by shifted black square patterns. (c) Herman grid illusion related to cognitive issues, i.e.
visual elements are generated where no elements are. (d) Müller-Lyer illusion related to
length, caused by extra visuals like arrow heads pointing in opposite directions [274]. Further
well-known effects are the spinning dancer illusion related to movement, caused by missing
reference points which seem to change the direction of movement, the Ponzo illusion related
to depth, caused by the environment and additional effect with denser becoming parallel lines
in the background like a railway track [274], or the checker shadow illusion related to color
and the surrounding colors [228].
for the same numeric value. This effect could happen in bubble treemaps or
node-link diagrams using circular shapes for the nodes with node weights
visually encoded by node size. Illustrating parallelism (b) could become a
problem for process visualization in which two processes are depicted as
running in parallel. Seeing patterns where no patterns exist (c) can lead to
misinterpretations of data and a longer response time due to visual elements
that have to be checked although they do not exist. Also the visual attention
might be misled which could be investigated by an eye tracking study. If
the length visual variable (d) is chosen to depict quantities, for example in
a histogram, those lines should not be attached by visual extras like arrow
heads to avoid confusion and wrong visual results.
2.5 Design Principles 73
Further visual illusions can be found that are worth mentioning. For
example, for a dynamic dataset it would be a disaster if the movement
effect was interpreted in several ways by different user groups or even the
same user. If depth is used to indicate older time steps or visual elements
not in focus we should be careful with comparison tasks that might lead
to wrong conclusions. Color is frequently used in visualizations to encode
a variety of data variables. If those are applied in certain environments,
typically combined with other colors, this might lead to an effect that causes
misinterpretations when judging and comparing the values of the colors,
hence the underlying data could be misinterpreted.
3
Visual Analytics
75
76 Visual Analytics
Figure 3.1 A quote by Albert Einstein or Leo Cherne describes the general ingredients
of visual analytics: “Computers are incredibly fast, accurate, and stupid. Human beings are
incredibly slow, inaccurate, and brilliant. Together they are powerful beyond imagination”.
cases the human analyst has to guide the process of combining the machine
with the provided algorithms and the implemented interactive visualizations.
Hence, the humans with their tasks at hand seem to be the major ingredient
in the visual analytics system. Those humans-in-the-loop, or sometimes
described as “human-is-the-loop” [174], can serve as participants in user
evaluations, with and without eye tracking, recording standard error rates,
response times, and also visual attention strategies recorded by more and
more advanced eye tracking systems [161, 235]. This kind of data about
human behavior has great value for understanding how and if visual analytics
systems function as expected or not. The real value becomes apparent when
visual analytics is applied again to the recorded evaluation data, supporting
hypotheses building, confirming, rejecting, or refining by taking into account
analytical approaches as well as the perceptual strengths of the humans’
visual systems. For visual analytics systems applicable to eye movement data
this can lead to a dynamic visual analytics system since it allows the recording
and analysis of the data in incremental and iterative processes.
key role. All of those interplays, benefits, and synergy effects have been
and are still developed by the visual analytics community, for example, by
communicating ideas at popular conferences such as the IEEE VIS with its
well-accepted VAST challenge. However, evaluation of visual analytics [306,
307], with and without eye tracking, still remains a challenging issue, not
only because all of those areas play a key role in the field, and this has made
it what it has become.
allowing to get web material that is of particular interest for the users. It seems
that the term visual analytics has already been used before the famous book
publication by Thomas and Cook [495], but what visual analytics actually
really means has not been clarified in the beginning.
Nowadays, many research fields are included to make visual analytics
successful and applicable to a variety of application examples. In their paper,
Wong and Thomas [529] stated that “visual analytics will not likely become
a separate field of study, but you will see its influence and growth within
many existing areas, conferences, and publications.” After several years of
development we have to admit that it has indeed become a separate field
of study, including more and more other fields, all giving benefit in their
role to find insights in data. Hence, visual analytics can only survive if
interdisciplinary research is done by several experienced scientists all over
the world while their research outputs have to be presented and shared at
famous international conferences like the IEEE VIS.
Figure 3.2 Data-related concepts may happen at three stages in a visual analytics system
in the form of preparing, checking and deriving, and advanced operations. Humans and
computers play different roles in these stages and are involved to varying extents.
checked for relevance concerning the tasks at hand. Data has to be annotated
or even de-annotated for anonymization for example, which might also be
done in rare cases by humans due to the lack of deriving the right computer-
supported semantics. Exploring the correct format as well as the data types is
as important as the linking between several data sources in case the linking is
not explicitly given. In all of these steps the human plays a crucial role while
the computer also supports most of the steps.
• Data collection and acquisition. Data can be collected in various ways.
If the user is involved, this typically happens in a hand-written form
with pencil and paper or directly stored in an electronic form, when the
situation allows it. However, to make the data explorable by a visual
analytics system it has to be brought into a computer-readable electronic
format. Hand writing can be read by a computer, for example by a
machine learning algorithm [366]. Normally, large data sources stem
from device-supported recordings (like eye trackers, cameras, sensors,
and so on) or dynamic and continuous data by simulations.
• Organization and relevance. In the case that a dataset consists of
several heterogeneous data sources, those must be ordered by relevance.
This decision is important to start with the most urgent data source
first, for example in situations in which a quick response is required.
The remaining data sources are read one after the other as soon as
computation resources are available. Such an organization is also of
interest if a primary data source has to be attached with additional
attributes from secondary, tertiary, and further data sources.
3.1 Key Concepts 81
(a) (b)
Figure 3.3 A node-link diagram in the field of graph visualization. (a) The relational data
without clustering, just randomly placed nodes. (b) Computing a clustering of the same data
as in (a) and, based on that, using a graph layout that takes into account the node clusters,
encoded by spatial distances of the nodes.
time periods and when the processing is ready, this data can be interactively
visualized and explored, not weakening the computing power of the visual
analytics system. This separation strategy guarantees a user-friendly tool
experience; no-one wants to wait for a long time for a data analysis result
while wasting valuable time. A data processing step should be as independent
as possible from the rest of the visual analytics system, in case the system and
the tasks at hand allow such a scenario.
Figure 3.4 Visual analytics is an interdisciplinary field that makes use of research disciplines
involving the computer, the humans, and also human–computer interaction (HCI).
there are too many to mention all of them here, but we will keep an eye
on the most prominent ones which are also involved to the largest extent. For
the algorithmic data processing we have to mention fields like data mining
and knowledge discovery in databases (KDD). Also machine learning,
deep learning, artificial intelligence (AI), or disciplines like explainable
AI build the basis for many data analysis techniques as well as statistics
and mathematics, in which the human user is typically not involved a lot.
Data management is important because of diverse data sources, and is also
available on the web existing in various and diverse research fields and
applications.
From the human perspective, fields like human–computer interaction
including individual and collaborative interaction, also over the web, play
a crucial role for more and more data analysts spread all over the world
with varying expertise. Cognition, perception, and psychology are important
for enhancing and accelerating the decision-making process, in particular,
if visualization is incorporated in the data analytics process, for example
90 Visual Analytics
in cases where large data like time-varying high volumes or streams are
visualized in a scalable form with a great overview first. Pattern recognition is
used for starting further exploration processes, to build new hypotheses about
the data, and to dig deeper in the already achieved insights. Presentation and
dissemination of the results is typically mostly the job of the humans although
computers are required to facilitate these issues. Evaluation is an increasingly
needed research area that provides insights into the user experiences and can
show design flaws and drawbacks that help to enhance the system. This brings
into play novel problems like ethics and privacy, creating a new kind of data
worth including in the analysis.
The field is still developing and has not reached the end of its progress.
As long as datasets are generated and recorded, visual analytics will play
a key role in order to process, analyze, and visualize this data with respect
to the tasks of the human analysts. However, due to the steady progress
in hardware and software technologies, visual analytics is also subject to
a steady change to keep pace with more and more challenging problems.
In particular, fields like machine learning, deep learning, or explainable
artificial intelligence [123] bring new ideas and emerging topics into play,
demanding the power of many involved aspects of this interdisciplinary field.
Taking into account the opinions, real-time decisions which are difficult for
visualization alone, and power of various users, maybe even in a collaborative
manner, is a suitable concept but requires scalability in terms of client–server
software architectures to reach out to the experts and non-experts in the world.
Visual analytics on smartphones, although the display is small, could be of
potential support, at least to access the people. Synchronizing, merging, and
summarizing the inputs and outputs of the users is another challenge. Such
apps would bring even more privacy issues into the field, a challenge that
cannot be neglected in the future. Users could suffer from cognitive overload
and hence, the problem might be split into parts and distributed among several
people to reduce this cognitive burden for the individual person.
Algorithmic and visual scalability are not the only challenging problems,
but also the general aspect of finding out if a visual analytics system is
useful at all or for which tasks it is of particular use. Evaluation is a huge
topic for the future development of such systems, specifically if millions
of people could be included in the evaluation process. The number of
people has to be large because visual analytics systems contain a variety
of functions, algorithmic approaches as well as visual depictions in lots of
parameter settings supported by interactions. Setting up user studies and
recording the performance or visual attention data is not the biggest challenge
3.2 Visual Analytics Pipeline 91
here, but it is more the effective and efficient analysis of the recorded data
with the goal to find or predict insights that help to improve the visual
analytics system, maybe in real-time. Further, automatic adaptations based
on users’ eye movement behavior or general behavior like body movements
and gestures, spoken words, interactions with the system but also between
the users, facial expressions, and many more could be steps into the right
direction. An exploration and interaction history, reflecting the steps and
stages taken during an analysis process, which is not found in many visual
analytics systems, can be of great support to achieve a faster way to find
insights, specifically if users have to jump to and fro between earlier and later
states of the system many times.
Figure 3.5 The visual analytics pipeline illustrates how data is transformed into patterns,
correlations, or rules that can be regarded as tables filled with values or visual depictions. The
users can adapt visual variables and refine parameters guided by their hypotheses and tasks,
hopefully generating new insights that can be used to modify the data under exploration.
These transformations are defined by the users with their tasks in mind,
hence due to the fact that the users can change their intentions quickly
it is important that the algorithms run fast and adapt the outputs to the
users’ needs. This is actually the stage in which the responsiveness of
the visual analytics system in terms of interactivity is based on. A poorly
running algorithm cannot be mitigated by a fast interaction technique. The
interactions do not only depend on the users’ intentions but also on the
algorithm performances. When working with algorithms it is crucial to have
profound knowledge about data structures on which the algorithms are based
and algorithmic runtime issues like NP-hardness [195]. In some situations,
the algorithmic problem is so complex that we cannot expect an optimal
solution in a reasonable time, hence a good heuristic approach is needed.
This is actually the point where we need experts in the field of data structures
and algorithms, typically needed to enhance the runtime and, consequently,
the interactivity of the system. In the cases in which we really need an optimal
solution for such a complex problem the users might be warned that a result
is not expected for an indefinitely long time. Showing a progress bar in such
a scenario is also not possible in most of the situations because we cannot
3.2 Visual Analytics Pipeline 93
predict how long the computation will take, i.e. we do not know after which
time period the one hundred percent is reached.
In some scenarios it would even be a good idea to provide insights
into the algorithm, which means showing the step-by-step iterations of an
algorithm and how it processes the data. This would help to understand how
long it might take until the algorithm terminates but, further, it would give
insights into why an algorithm produced the wrong results or suffers from bad
performance. Although such an approach would be suitable for the design of a
visual analytics system, it is less useful for an end user who plans to just apply
the system to find insights into the data based on the tasks and hypotheses.
However, the algorithmic issues could be reported to the developer of the
system while the users let the tool run without recognizing that such extra data
is stored and transmitted. By such a strategy the system might be evaluated
from an algorithmic perspective, but the end users’ feedback could also be
included and linked to the algorithms under different settings.
effortlessly detect insights, in case the chosen visual metaphor is the right
one for the task at hand. Hence, a visual depiction of the data is of particular
importance to support the viewers and to guide them in the right direction as
well as accelerate the insight detection process. This is, in particular, useful
for applications in which a quick answer to the problems at hand is needed,
like evacuation planning, epidemics, or preventing terror attacks [495].
A lot of experience is required to design a good visualization as well
as to read it and to derive patterns, correlations, and rules. Moreover, it
should be investigated whether visualization is really needed to solve the
given task or if a pure algorithmic solution is strong enough to provide
the right answers. Visualization is the means of choice if an algorithm
cannot be clearly specified in terms of parameters and details about the
computation steps. An example is the min–max search in a set consisting
of quantities. This task can be solved on a sole algorithmic basis without
asking a diagram for support. However, if we search for a pattern which is
very vaguely defined, also depending on the dataset and domain, we might
choose a visualization. But to choose that in the right way we need profound
knowledge about the data types and structures involved in the dataset as well
as the domain and environment from which it stems, also the user tasks are
deciding factors in picking the right visual metaphor and right individual
visualization techniques. In summary, applying visualization for a dataset
to detect patterns, correlations, and rules is a challenging task, not only on
the design level but also on the interpretation level, typically involving non-
experts in visualization. This is the step in which user evaluation is important
to understand if the designer was successful or not in supporting the pattern
finding task which is interesting for the user.
A pattern is a user-defined order or structure in a dataset that follows a
certain behavior, shape, model, outline, or template. It is something that is
the opposite of random. A pattern carries some kind of user-defined meaning.
In some cases it can be clearly specified in terms of parameters, but in most
cases the users are able to define it as a pattern although they cannot specify it
very clearly which makes it hard to be processed by an algorithm. Moreover,
in some data situations it was not clear right from the beginning that a pattern
existed in the data, hence an algorithm cannot be specified which refers to
the “seeing the unseen” quote. If a pattern can be identified, also outliers and
anomalies can be present which are points that do not fall into the pattern
shape, but behave somehow differently. Outliers and anomalies can only
exist if a pattern exists. A powerful visualization for depicting patterns is
an adjacency matrix, if it is ordered in a meaningful way. Figure 3.6 shows
3.2 Visual Analytics Pipeline 95
(a) (b)
Figure 3.6 (a) Comparing the participants’ scanpaths from an eye tracking study can
generate pairwise similarity values shown in an adjacency matrix. (b) Applying a matrix
reordering technique immediately shows a pattern in the matrix which is difficult to find by
the pure textual values given in a table or 2D array [300].
two examples for the same dataset, an unordered one and an ordered one.
The ordered matrix immediately reflects patterns in form of blocks along the
diagonal.
A correlation is defined as two or more variables measured for the same
observations standing in a certain related behavior. This behavior can be
the same, similar, or one can be completely the opposite of other ones. An
example would be a bivariate dataset, i.e. containing two variables for each
observation, in which the larger values of variable A correspond to the larger
ones in variable B and the lower ones of A correspond to lower ones of
B, building a positive correlation. If the larger ones of A correspond to the
smaller ones of B and vice versa we call this effect a negative correlation. If
those values are stored in a table with two columns, an ordering of one column
can help us to see the impact of this ordering on the second one. Scrolling
down the columns can still be a solution to identify the correlation behavior
but in many cases it is not that easy to identify in the textual representation.
A visualization can help to rapidly judge which kind of correlation (apart
from the standard positive and negative ones) exists. Scatterplots for bivariate
data as well as scatterplot matrices and parallel coordinates plots [255] for
multivariate data are standard and well-researched depictions for this kind of
data focusing on visual correlation detection. Figure 3.7 shows an example for
multivariate metric data from an eye tracking study [299]. The vertical axes
encode the metric scales from top (large values) to bottom (small values)
96 Visual Analytics
Figure 3.7 A parallel coordinates plot (PCP) for showing positive and negative correlations
between pairs of metric attributes derived from eye movement data for selected study
participants [299]. Axis filters are indicated to reduce the number of polylines. Image provided
by Ayush Kumar.
and the polylines in between depict the values for each observation, in this
case an eye tracking study participant. We can detect the positive correlations
in the nearly horizontal more or less parallel running lines and the negative
correlations in the crossing line patterns between the axes.
If data mining is applied we can generate more advanced rules that can
exist in at least two major types denoted by association and sequence rules.
Depending on the number of elements involved they can be binary or n-ary.
An association rule expresses if two or more data elements are related at the
same time to a certain extent, typically described in a percentage value or
in natural numbers. This depends on the fact of whether the confidence or
the support of a rule is considered. The support expresses the total number
of occurrences of an element tuple in a rule while the confidence expresses
the relative number of an occurrence of an element tuple to all occurrences
of an element in that tuple. A sequence rule, on the other hand, contains
a temporal aspect, describing an antecedent, the condition before, and a
consequent, the condition afterwards. Also sequence rules exist with a certain
confidence, i.e. probability, as well as a support value, even in multiple stages,
if the sequence rule consists of several antecedents and consequents. Each
following consequent typically lowers the probability because of the more
restrictions due to the higher number of antecedents. Figure 3.8 shows an
example from eye movement data [63] for which data mining rules have been
generated and visualized. A typical visual aggregation for sequence rules can
be achieved by considering common prefixes and put them together into a
rule hierarchy, just like some kind of decision tree.
3.2 Visual Analytics Pipeline 97
(a) (b)
Figure 3.8 (a) n-ary association rules (b) and n-ary sequence rules generated from eye
movement data can express which general relations exist in eye movement data [63].
It may be noted that apart from patterns outliers and anomalies can also
exist, but those can only be detected if a pattern is known, i.e. the normal
case from which an outlier can be distinguished. A similar effect holds for
countertrend patterns that can only be detected if we know the trend pattern in
the data. For example, if one time-series shows a growing behavior, a second
one shows a decreasing one. However, it always depends on the perspective of
the user what is defined as the trend and the countertrend. A trend is typically
found in time-dependent data in which we can compare values over time.
This means we identify a certain function that describes or models the trend
behavior as well as possible.
Figure 3.9 Confirming or rejecting a given hypothesis in a simple bar chart can generate a
lot of simple tasks which might be recognized if the eye movements of observers are recorded
and overplotted on the bar chart stimulus in the form of a gaze plot [203].
on the scale (to check for 8), and finally, read the labels. This visual stimulus
also shows some extras, one is definitely the fact that not all bars are visually
attended, a fact that might come from the peripheral vision, meaning in some
situations we do not have to focus the eyes on a visual object to judge if it is
relevant for a task to confirm or reject a given hypothesis.
We have to admit that the bar chart example is not needed to answer the
given hypothesis, this could be solved by an algorithm while the labels could
also be checked by just searching for them in the dataset. However, there
are much more complex hypotheses for which visualization is required. For
example, we might state that “there exists a periodic dynamic pattern in the
data”. This hypothesis could be algorithmically solved if we knew the period
in the time-varying data, but this period might even change over time. Hence,
a visual depiction would help and the perceptual abilities allowing fast pattern
recognition would do the rest for us.
The examples above illustrate that a simple hypothesis requires a certain
number of simple tasks that have to be combined in order to find a way to
confirm or reject this hypothesis. In general there are various simple low-level
tasks that can be combined in task categories, each describing the common
procedure that is needed to solve each individual task contained in it. In a user
study, with and without eye tracking, the participants are typically confronted
by one or several of those tasks, simple ones or more complex ones. If the task
is too complex the study participants start to subdivide the complex task into
simpler tasks automatically. This subdivision can be identified if the visual
attention behavior in form of eye movement was recorded and the analysis
approach was sufficient to identify the task splitting strategy. In the following
an overview is given about certain task categories without explicitly stating
that the list is complete (see Figure 3.10). The tasks can be applied to both,
textual representations in the form of tables and values, but also to visual
depictions of data, probably with varying task performance in the form of
task response times and even error rates, typical for user studies reflected
in the dependent variables. Yarbus [539] has shown that the task even has
an impact on the eye movements, hence it would be interesting to analyze
eye movement patterns to identify the current user task in a visual analytics
system, maybe to better guide a user.
• Search tasks. Locating textual or visual objects of interest builds the
basis for nearly every other task. Search tasks can be quite time-
consuming if the objects we are looking for are not quickly recognized,
for example outliers should be visually highlighted and hence, be made
pre-attentively detectable [219, 500].
100 Visual Analytics
Figure 3.11 Changing the requirements for an algorithm can modify its output, which is seen
in a visual result. In this case the layout algorithm for a generalized Pythagoras tree [31] for
hierarchy visualization (top row) is changed to a force-directed one (bottom row) that creates
a representation which is free of overlaps [363].
wait, valuable time that might be lost in which we might have done something
else. If an urgent and timely solution is required, for example in an application
scenario in which the users are under an extreme time pressure for making a
decision, such an effect is not desirable and should be avoided when possible.
However, some algorithms do not allow for a quick solution, even on the
fastest computer [195], hence the user should be warned in such a situation.
Most modifications, either to the parameters of the algorithms or to the
visual variables, have an impact on the other side, i.e. algorithms influence
visualizations and vice versa. This means, for example, if an algorithm setting
is changed, the new output will have an impact on the visual result, for
example a refinement for the layout or arrangement of the visual objects.
On the other hand, changing a visual variable might request an algorithm to
compute a more efficient structure of a set of elements. This bidirectional
interplay between algorithms and visualizations is important for a visual
analytics system but also requires that the human observers are able to keep
track of what has been changed and what stays the same or similar. This
effect brings into play the mental map [356] which is a crucial aspect for
keeping a system with visual and algorithmic ingredients usable and user-
friendly. Such refinements and adaptations bring typical parameters in user
experiments to check their impact on the user performance, but also on
algorithmic performance without including humans. For eye tracking it is of
particular interest that the visual analytics system can provide various types
of visual stimuli to investigate the impact of parameter changes on them,
and, further, on the interpretation of the users combined with the tasks. Only
with these insights are we able to find out for which setting a certain design
104 Visual Analytics
flaw occurs that needs to be removed. Moreover, with eye tracking we can
record where (space) and when (time) the design problem appeared and who
(participant) had bad experiences with it, reflected in the visual attention
behavior.
The evaluation of such refinements and adaptations does not only depend
on the shown visual stimuli but on even more factors. Interactions are
supported by various input as well as output devices, can be done by
individual users or in a collaborative manner, and the users themselves with
their experience levels as well as tasks in mind build a crucial ingredient.
Moreover, the eye can itself be used as interaction means, for example
as gaze-assisted interaction [439], a concept that already records the eye
movements during interacting with a visual analytics system.
semantics, visually describing how a given task might be solved. If the insight
is not enough at any stage the user does another iteration for which the visual
attention behavior could be compared with the previous one to analyze which
chunk of information was missing in the previous iteration. It can be evaluated
if insights are as expected, good enough, how fast people came to them, if they
made misinterpretations, and what they visually attended. Insight is much
related to user confidence, meaning the more insight the users get, the more
confident they are and the fewer eye movements are made from iteration to
iteration, but as long as this is not properly analyzed the effect just serves
as a hypothesis about the eye movement data. In any case the user getting
frustrated while looking for more and more insights should be avoided. This
effect might be detectable in the eye movement behavior, maybe by more
and more chaotic visual attention patterns. On the positive side, the faster the
insight is generated, with a minimum number of eye fixations, is a good sign
for a well-designed system.
A maximum of insight finally leads to knowledge about a certain data
scenario, aspect, or situation to allow making a decision. For example, in
urgent cases for timely decisions with a lot of stress for the user, the eye
movements might vary a lot [539] compared to a totally relaxed situation for
which the rapid insight detection is not required but for which we have all the
time in the world. Also for a real-time visual analysis in which the user cannot
see the dynamic stimulus easily again and again, like in a replay, without
losing real-time information, we could analyze the recorded eye tracking data
to identify where the design flaws occurred for generating insights, hence
real-time situations are challenging to evaluate by eye tracking. This is also
problematic for a feedback loop allowing users to step back and adjust or
adapt certain parameters, with the challenging issue that the real-time analysis
is running at the same time as we wish to step back. For a non-real-time
situation the feedback loop is a powerful concept since it allows us to go back
to the data stage, but also to stay in the pattern, correlation, or rule generation
stage, and refine and adapt it until insight is generated again. Visual analytics
provides many opportunities to analyze data and to finally get knowledge
about the data under exploration. To sum up, those insights are important to
present, share, and disseminate to and with an interested audience.
layout for the users’ tasks at hand, which gets even more challenging for a
dynamic graph, i.e. one that changes over time for which the dynamic stability
has to be taken into account to support the user as good as possible to preserve
the mental map when using an animation to explore the dynamic data for
trends, countertrends, or anomalies.
still have the option to try another kind of algorithm that might have a higher
runtime complexity but successfully generates the patterns we wish for.
For algorithms from the class of online problems, we mostly have one
chance to watch the data, because it is shown in real-time, hence the correct
parameter setting has to be chosen in the beginning. This could be based on
learning from previously seen scenarios or the data must be recorded and the
important time periods will be observed in an offline setting after the real-
time application has ended or even simultaneously, if the time permits. An
offline inspection of the online data provides more time to explore and adapt
the parameters to the users’ needs, but in cases we have to react in real-time
due to timely reactions, a visual analytics system must take into account such
a complex scenario, even the combination of online and offline algorithms.
It should even be possible to adapt parameters to modify the phase of an
algorithm during runtime. For example, if visual output is given, even in form
of a simple statistical value like a threshold number, we could initiate the
algorithm to run in a new phase, for example from searching for an optimum
to just a local maximum. This makes sense if the algorithm runs into a time-
consuming routine with bad performance.
g(n)
o(f ) := {g : lim = 0}
n−→∞ f (n)
f (n)
ω(f ) := {g : lim = 0}.
n−→∞ g(n)
(a) (b)
Figure 3.13 Runtime performance chart for two algorithms generating visualizations. (a) A
word cloud is generated for differently large dataset sizes, resulting in a linear-like runtime.
(b) A pedigree tree is generated based on more and more people involved, standing in a family
relationship, resulting in some exponential-like runtime.
different way. In such a scenario we could start the visual analytics system
and for every computation that has to be made we could store the time
it takes from starting the computation until the result is produced. This
process could be done for several instances of the algorithm responsible for a
certain computation by increasing the size of an input parameter, for example,
the dataset size in terms of number of elements to be processed. Even
an application on a different machine or a different environment can have
an impact on the runtime performance. Plotting the measured performance
values (see Figure 3.13), i.e. the time in this case, on the y-axis of a coordinate
system and the size of the varied parameter, in this case the number of the
data elements, gives a visualization of the runtime function. Having some
experience with such mathematical functions and what they look like when
plotted can give us an impression of which complexity class the algorithm
falls in; however, it may be noted that this is not a mathematical proof, it
is just a hint of what the runtime of an algorithm might be, which cannot be
modeled in terms of input parameters and a function depending on these input
parameters.
A performance evaluation might even be based on the memory
consumption in addition to the runtimes. Negatively, if we use Landau
symbols to express runtimes or memory consumption we only get the total
aggregated performance of an algorithm after it has run. In a strange scenario,
an algorithm might run very efficiently, maybe in O(n) time for most of
the processing steps, but negatively, a few processing steps may take more
time than expected, which would result in a much longer runtime. The
aggregated performance is not able to explain this phenomenon, which makes
it absolutely necessary to explore the internal step-by-step processes of an
114 Visual Analytics
(a)
(b)
Figure 3.14 During an algorithm execution, the runtimes (a) as well as the memory
consumption (b) might differ from iteration to iteration.
3.3 Challenges of Algorithmic Concepts 115
Figure 3.15 A time-to-space mapping of the vertices and edges processed by a Dijkstra
algorithm trying to find the shortest path in a network is visually represented in a bipartite
layout [75]. The time axis runs from left to right.
system it would be an extra insight if the users could choose to have a look
into an algorithm and see how it works step-by-step (see Figure 3.15). Such
an option would not be used by domain experts who would just explore their
data, but by the designer of the system to understand the design flaws, not only
on a visual basis but also on an algorithmic one, but supported by algorithm
visualizations, typically as animated sequences consisting of the individual
steps [59, 479, 480].
Before we can visualize the instances over time that an algorithm
generates we have to access and store all the information that is required
to get a suitable visual depiction of such a dynamic dataset in order to use
it for insight detection. The tracking of the variables in use as well as the
internal step-by-step states of the transformed data structure is challenging
but can be of great help if visualized in an efficient way. We could even argue
that the algorithms incorporated in a primary visual analytics system might
be so complicated that another secondary visual analytics system is needed to
analyze their internal workings for detecting design flaws in the algorithms of
the primary visual analytics system. This procedure would make the primary
visual analytics system a dynamic visual analytics system since the generated
insights can be used to improve the algorithms, their runtimes as well as their
visual representations, if eye tracking is applied as well.
To show the step-by-step execution of an algorithm [282] we have two
major options which are denoted by time-to-time or time-to-space mapping.
A time-to-time mapping also contains the concept of animation [505], which
maps each time instance in a dynamic dataset to physical time, typically as
a smooth animation with smooth transitions from one step to the next one
116 Visual Analytics
to distinguish it from just showing individual time snapshots one after the
other. In contrast, a time-to-space mapping is a static representation of the
dynamics of the algorithm. The benefit of such a static representation is that
it shows several instances at a time in the same display space which makes
the dynamic data comparable over time. This is difficult, even impossible, for
animated sequences. We have to stop and replay, but still comparison tasks are
hard to solve successfully. A negative issue of time-to-space mappings is the
fact that the display space has to be used with care to show as many time steps
as possible, otherwise the individual time instances with their corresponding
visualizations might get too small to be inspected visually. A concept that
is between time-to-time and time-to-space mappings, some kind of hybrid
approach, is rapid serial visual presentation (RSVP) [477]. A variant of it
is called dynamic RSVP that shows a sequence of individual time steps as a
time-to-space mapping and animates a sliding time window over the sequence
to show the progress over time. This always gives an observer the chance to
inspect a sub-sequence of a well-defined temporal length.
3.4 Applications
Visual analytics has existed for quite some time now which is the reason why
it is applied in many application fields focusing on a variety of data analytics
problems. The repertoire of examples for such tools is increasing day-by-day,
the more complex ones focusing on heterogeneous data sources accessible
online, while the most successful ones are typically published at renowned
conferences or workshops or in journals and books. Some of the researched
solutions even made it from academia to industry as a powerful concept to
analyze specific datasets; however, for most of them, domain knowledge as
well as some expertise in algorithms and interactive visualization techniques
are required. From the various application examples we will just showcase a
small selection for illustrative purposes.
The applications can also be distinguished by the size and the complexity
of data they can handle, the types of tasks, the target users, like decision
makers, who might have to react in real-time, or whether they have been
designed for the expert or non-expert in a certain domain. Some of the
prominent application fields are astronomy or astrophysics with continuously
changing data streams full of noise, seismology with geographic information
measured over several attributes, weather, climate, and meteorology focusing
on predicting scenarios based on sensor and satellite data, security containing
people networks and their interactions based on social media and global
3.4 Applications 117
Figure 3.16 A flight traffic dataset taking into account temporal clusters while a bipartite
splatted vertically ordered layout is chosen to reflect static and dynamic patterns in the time-
varying graphs [2]. Image provided by Moataz Abdelaal.
118 Visual Analytics
Figure 3.17 The graphical user interface of the pathology visual analytics tool with the
image viewer, the image overview, a gallery with thumbnail images, a textual input to make
reports, a scatterplot for showing correlations of bivariate data, and a view on sequential
diagnostic data [134]. Image provided by Alberto Corvo.
3.4 Applications 119
or Utopian future wishes, visual analytics can at least support several, if not
all, data analysis tasks for clinical researchers, given the fact that the data is
available and ready to be used [135].
Typical tasks that are of interest come in the form of aggregating the data
based on patient groups and their symptoms to derive a general kind of rule
that holds if a certain subset of symptoms occurs for a special patient group.
Based on such generated rules a prediction or therapy could be suggested
while the new generated data and the reliability of the rule could be stored
in the database with the goal of improving the accuracy of the set of rules.
Data mining could be an important discipline here which shows again that
visual analytics is some kind of interdisciplinary approach that has to take
into account various technologies to be as powerful as possible. Insights
from applied algorithms could be useful to find abnormalities in the data,
facilitate the identification of a characterization that on the other hand helps
to better group and categorize the data on several attributes and criteria, or,
in general, information could be derived that is not directly visually observed
from shown medical images for example. However, one problem this field
has to deal with is of an ethical nature, i.e. the requirement for respecting
individual human rights and privacy aspects. These challenges make data
analysis and, in particular, the dissemination and sharing of the results to
a larger audience difficult, although it would be a great benefit to find more
reliable and accurate therapies given the fact that all of the patients’ data is
publicly available. The question is if the benefits outweigh the drawbacks and
misuses of the data.
Figure 3.18 EventPad [106] is based on a graphical user interface with several interactively
linked views to support data analysts at specific tasks to explore malware activities. Image
provided by Bram Cappers.
Figure 3.19 Visual analytics supports several views on the video data [232]: time navigation,
video watching, snapshot sequence view, audio augmentation, statistical plots, graph views,
schematic summaries, and filter graphs. Image provided by Benjamin Höferlin.
semantic information or visual patterns that might carry some meaning (see
Figure 3.19). In video visual analytics [233, 234] we try to find reliable and
fast answers to the tasks at hand, in particular, if the data stems from video
surveillance to investigate criminal cases. Those tasks might be to compare
video sequences to identify people who act in different scenes at different
times, find analysis-relevant time periods in a video to reduce the time for
browsing the video, or to detect certain well- or partially-known patterns,
also known as weighted browsing. Visual analytics also takes into account
additional information to augment and annotate a video for the observer, for
example, based on object detection and tracking. Moreover, further derived
information from the video sequence like color distributions might be of
interest to apply some kind of temporal clustering or to faster compare several
video sequences.
Watching the video to identify certain patterns that bring us to the
solutions of analysis tasks would be one way but it requires to watch each
video completely. For data that is produced from video surveillance systems
this strategy is too time-consuming and would not quickly enough lead to
a desired goal. Making a quick decision based on the gained insights is as
important as to understand the relations between the environments, objects,
and actors in a scene recorded as a video. Automatic analyses alone are a
powerful approach to quickly examine the time-varying data but, negatively,
the semantic information cannot be reliably extracted and judged by a pure
machine-based approach. This is the point at which visual analytics comes
into play since it combines the power of the machines with the perceptual
abilities of the human observers being able to recognize patterns and to
explore them in context to a semantic meaning that the machine is not able to
derive.
122 Visual Analytics
Figure 3.20 GazeStripes: visual analytics of visual attention behavior after several people
have watched videos [309]. Images provided by Kuno Kurzhals.
visual attention data as well as the data stemming from the gaze-assisted
interaction. Another challenge in this line of research is whether the recorded
eye movement data can be reliably analyzed to enhance a visual analytics
system based on users’ visual attention input, making it to a dynamic visual
analytics system.
4
User Evaluation
125
126 User Evaluation
Figure 4.1 The most important ingredients in a user study are the participants, the study
type with the independent, confounding, and dependent variables, and the results in the form
of statistics and visual depictions.
requires some experience about the repertoire of existing eye tracking data
visualizations [47]. However, both kinds of evaluations, performance and
user evaluation, provide a wealth of insights for enhancing a visual analytics
system, algorithmically or perceptually.
Evaluating visual analytics systems and analyzing the recorded
performance data can have several benefits, for example, to obtain measured
values instead of subjective feedback and personal judgments, but in case that
spatio-temporal eye movement data is incorporated in the analysis process,
we have to tackle many challenges before insights about the user behavior
can be detected. One benefit is definitely that design flaws can be identified
over space and time, even for certain specific participant groups, which leads
to an improvement of the quality of the visual analytics system, maybe
only for certain groups of users. This quality can be measured during the
development or after the visual analytics product is finished, making it a
problem- or technique-driven evaluation [450]. In the case that there are
serious design flaws which lead to a degradation of performance at some
task [426], the system might be redesigned in some way and evaluated again
to check whether there is any kind of improvement. Even before starting
the design and implementation phase, a user study, in this case more in
the form of interviews, could help to guide the development process in the
right direction. If several system variants are available and there is no clarity
as to which one is better for certain tasks and user group, evaluation can
help as a means to get a qualitative or quantitative comparison between the
variants [397]. Moreover, not only can the visual analytics system be the focus
of an evaluation, but also the users, allowing categorization into user groups
4.1 Study Types 127
records the data over space and time accurately. In addition, it should be
checked whether the participants who are not familiar with eye tracking can
effortlessly work with the new technology without being too stressed and
feeling uncomfortable. Eye tracking studies are very different from regular
studies without eye tracking.
important aspects that have been said during the experiment. Think- or talk-
aloud studies might have an impact on the participant’s behavior as well as
writing down notes. This is even more problematic in remote eye tracking
studies in which the participants should not move around too much to avoid
calibration errors, and hence inaccurate scanpath recordings.
For eye tracking technologies it makes no difference if a quantitative or
qualitative study type is used. The eye movement data can serve as additional
input for either enriching the dependent variable response time with temporal
visual attention information to identify the causes of response time variations
between stimuli or when the time is wasted, why an error has been made,
and which visual region in a stimulus might be problematic and worth
enhancing. For a qualitative study we might enrich the verbal feedback by
visual attention data to understand why, where, and when a participant made
the comments. This gives a bit more insights from the cognitive processes
that are not recordable by eye trackers. However, linking eye movement data
with other data sources like qualitative feedback is a challenging task and
likely producing interpretation errors. For visual analytics, both, quantitative
as well as qualitative study types are useful. Quantitative studies can help to
compare simple stimuli like static diagrams to find out which one would lead
to better performance measures for a specific task at hand. Qualitative studies,
on the other hand, might be useful in more complex scenarios, even during
the design process, but the generated results are less exact due to the missing
quantitative values.
Figure 4.2 A comparison between Cartesian and radial diagrams in an uncontrolled user
study recruiting several hundred participants in an online experiment [150]. Image provided
by Stephan Diehl.
(see Figure 4.2), mostly being of a short duration of a few minutes to attract
as many people as possible, sometimes offering money or an additional gift
for an elected winner. If there seems to be a common pattern or strategy, this
typically shows an impact from an independent variable on a dependent one,
but still there is no guarantee for it, even if thousands of people show the same
behavior.
We might say that the more control is forced on a study the smaller the
population in the study. The question is, how we can put as much control as
possible in a study while at the same time recruiting a large number of study
participants which would allow us to increase the number of independent
variables to be varied? This limitation of controlled studies has an impact on
the visual stimuli, being rather static and trying to focus on comparing two or
more visuals based on the same task. This might result in an assumption that
one technique is better in terms of error rates and response times than another
one under certain circumstances. For uncontrolled studies the parameter
space can be larger, i.e. showing an interactive stimulus like a visual analytics
system and allowing many people to experiment with it, maybe asking for a
132 User Evaluation
might not be caused by the task applied to a visual stimulus but by other
influences, like the technology, which is another confounding variable in the
study. It is always a good advice to conduct the study with non-experts first,
and then refine the tasks, stimuli, and independent variables to reduce the
study complexity. As a second stage we can use the adapted and refined
version for the few experts that are available, hence we should not waste our
valuable resources on an ill-designed study. As another interesting insight to
explore the expertise of a participant, the scanpaths can be compared which
could reflect expertise changes over time, i.e. with the progress of the study,
but this insight could even be used to classify expert and non-expert users.
matter which type it is. This requires that the users-in-the-loop, i.e. the study
participants, have to be tested for a variety of aspects and properties before
running the study, even before thinking about the final design of the study.
Those human aspects definitely play a major role in the study setup, for
example, they might even serve as independent variables to measure their
impact on the dependent ones like error rates, response times, or visual
attention strategies recorded in an eye tracking experiment.
The human users can play different roles in a study, typically described by
the purpose of the study or the goal it is based on, like exploratory, predictive,
formative, or summative [12]. These goals might be used to build another
kind of study type classification in which the study participants are more
closely taken into account than in the study types discussed in this book
(Section 4.1). The individual users with their roles can have any kind of
background knowledge but before starting a study, or even before designing
it, the experimenters typically try to find the best possible way to recruit study
participants [171]. In the recruitment process there is some kind of trade-
off since the expertise of the participants stands in a negative correlation
behavior to the relationship between the experimenter and the participants
themselves. This means that the more expertise the population taking part in
a study has to solve a given task, the farther away this population is from
the personal environment of the experimenter. For traditional small-scale and
simple-task studies this trade-off might bring some kind of bias in the study
if this problem is not carefully taken into account.
(a) (b)
Figure 4.3 Examples from a visualization course at the Technical University of Eindhoven
educating students in eye tracking and visual analytics [70]. (a) A visual attention map with
contour lines. (b) An eye movement direction plot.
task and to understand the visual stimuli, like a visual analytics process,
the students might not be the right choice for a study population. To raise
them to a certain level of expertise they are educated, in the best case, in eye
tracking [71], and hence, trained during the lecture to finally serve as some
kind of experts in the study. After some training of about eight weeks the
students are able to design and implement their own interactive visualizations
for eye tracking data (see Figure 4.3), enriched by algorithmic concepts to
transform the data [70], showing a steep learning curve, which means they
obtained profound knowledge of this specific application domain that they
did not have before attending the course.
The training has an impact on the individual performance, possibly
introducing some kind of bias. It is important that the individual participants’
expertise after the training session is checked by test questions and a thorough
practice runthrough. All of the participants should be on a similar level
after some time to take part in the study. Randomization, replication, and
permutation are powerful concepts to average out certain differences in
the expertise and performance that might even vary during an experiment,
typically reflected in learning or fatigue effects. Measuring the expertise is
difficult but not impossible. It can be based on the accuracy of the answers
to, and response time to, test questions in the practice runthrough since only
the expertise relevant for the particular study is important, not the expertise
stated by a participant. People could even be excluded from a study based
on the level of expertise, but this should be done after the participant has
completed the study. Moreover, the rules for such an exclusion must be
4.2 Human Users 141
several senses have to be used in combination to make the best of the system.
But such diseases can be a great opportunity for user studies since they can
serve as a way to understand the difficulties this population group has in
daily life, hence evaluating a system for insights can be a useful strategy to
improve the daily living environment for these people. The biggest challenge,
however, is to get participants for a study related to such issues, also for
the age group of young children, maybe under the age of six. Ethical issues
also build a limitation in this research field and hinder the development of
appropriate tools and systems.
For eye tracking, age plays a crucial role, not only due to reasons of the
study setup like the sitting position’s height and distance to the monitor to
allow a smooth and comfortable calibration and tracking phase for a remote
eye tracking device. Moreover, also many instructions are dependent on the
age of the participants in an eye tracking study; infants must be controlled
a lot more than adults, while elderly people may need many more details
explained on the technological side. This is the case for controlled studies in a
lab, but even more for uncontrolled settings in the field. For visual analytics it
could make sense to include very young people, in case they serve as domain
experts, for example, for analyzing data stemming from a certain kid-like
environment like a kindergarten in which data about strategic game playing
is recorded. A study setting could make sense in which a nursery teacher
and a child watch such scenes on a monitor while the kid is eye tracked
and the teacher tries to figure out the visual attention behavior of a kid in
a specific scene. In this case eye tracking is required to produce another kind
of data to the given video stimulus data as well as visual analytics to let the
experimenter, here a nursery teacher, visually explore the data. For elderly
people we could find similar useful scenarios, focusing on understanding
the problems they might suffer from in their daily lives with the intention
of enhancing their situation based on the insights found by applying visual
analytics.
(a) (b)
Figure 4.4 (a) A Snellen chart [470] can help to identify visual acuity issues. (b) An
example plate of an Ishihara color perception test consisting of several pseudo-isochromatic
plates [258].
4.2 Human Users 145
advisable to respect the people involved in the study; this holds for the study
participants as well as for the experimenter who can be mistreated without
respect, for example, if people just take part for the reason of earning money.
Everybody should be treated similarly, no matter which gender, age, culture,
religion, and so on this person is associated with.
For the recorded data there could be the general problem that the results
are not convincing enough or not convincing at all. In these days with various
other research groups working in the same or similar area it is tempting to
manipulate the results in a way to obtain more convincing results. Such fraud
must be avoided, but it might be quite challenging to detect it. For this reason
it is desirable to share the recorded data to make it publicly available for the
research community to check for any inconsistencies in the analysis steps.
However, the data might already have been manipulated before the sharing,
which is hard to avoid. Some kind of mechanism would be beneficial that
directly shares the recorded data in its raw form which would minimize the
chance of modifying to get more convincing results. However, at least in the
visual analytics field, such an approach is hard to realize and might contradict
the goal of data protection and data privacy. At least, in many institutions
there is some kind of ethics committee that checks the rules before running
an experiment. Negatively, it can take quite a long time until permission is
given to start a study, i.e. an ethics committee can sometimes slow down
research, but it is definitely needed.
People tend to be more under pressure in an eye tracking study than in a
traditional study without eye tracking from the perspective of ethical reasons.
This is due to the fact that people might fear that they are more observed
because of the general opinion that the eyes tell more about a human than any
other organ. The eyes are some kind of window to the soul [246] and it is
said that emotional states can be recognized [332] by reading the eyes as well
as if somebody is telling the truth or lying. Ethics can be a problem in eye
tracking research [320] because it might be unclear for the ethics committee
to assess how invasive the procedure will be, if eye damage can occur, and
which additional body-worn sensors are used to record additional personal
and private data, which can cause a delay for the confirmation to start a study.
This delay is typically caused by missing information about the relatively
new technology, in particular when evaluating visual analytics systems. For
example, the types of visual stimuli and the way they are presented plays
a crucial role, also taking into account that people might have health issues
when exposed to a flickering stimulus, maybe causing epileptic problems.
People should be informed about the option to give up, even if they are forced
4.3 Study Design and Ingredients 147
to take part in the study with the money as some kind of pressure. Moreover,
the data recorded in an eye tracking study contains much more confidential
and privacy information than the data from a traditional user study without
eye tracking.
typically happen over time. The number of replications decides the coverage
of certain configurations of the independent variables. A pilot study is an
appropriate way to figure out if all the study parameters are well chosen or if
some have to be modified and adapted to guarantee a smoothly running actual
non-pilot user study. For eye tracking studies, the technology also has to be
included in the study design process in an adequate way. Actually, in most
cases, if we can run the study without eye tracking we can also run it with eye
tracking by just using suitable eye tracking technology, i.e. either remote or
wearable eye tracking systems. The downsides of adding eye tracking as an
opportunity to record additional performance measures in the form of visual
attention can be tested in a corresponding pilot study.
might argue that variant two is better than variant one of a visual analytics
system by investigating the judgments or personal opinions of the study
participants. However, we could also measure performance but in general
in such a complex visual stimulus and study setting it is hard to compare
performance due to the fact that the people start interacting and follow
different exploration strategies. This makes the recorded values incomparable
due to the fact that the concrete task is split into different subtasks, each
participant uses a different subtask organization which is one reason why
a hypothesis is typically not checked with raw numbers in a statistical
evaluation.
Eye tracking can give insights into the visual scanning strategies and
allows comparisons between the participants, even if they followed different
subtasks to respond to the main task. Hence, eye tracking is a suitable way
to confirm or reject hypotheses focusing on such viewing behavior, i.e. over
space and time. However, the hypotheses are difficult to statistically evaluate,
but visual analytics can serve as some kind of evaluation since it allows
analytical reasoning incorporating algorithms, interactive visualizations, and
the human user (Chapter 6). As a challenging aspect, the hypotheses in eye
tracking studies can be much more complex due to the spatio-temporal nature
of the recorded data, even combined with extra data sources [44]. Hypotheses
might be built that refer to the space and time dimension in the data at the
same time. Moreover, the semantics of a (dynamic) stimulus can be taken into
account, like getting some information in a certain region at a specific time
point that is applied later on in a different region in the dynamic stimulus.
Such a scenario brings challenging issues related to cognitive processing, for
which we cannot easily find answers.
(a) (b)
(c) (d)
Figure 4.5 The way a stimulus is presented and the degree of freedom of the participant’s
position has an impact on the study design and the instrumentation. (a) A static stimulus, like
a public transport map [372], inspected from a static position like sitting in front of a monitor.
Image provided by Robin Woods. (b) A dynamic stimulus, like the game playing behavior of
people recorded in a video [71], inspected from a static position. Image provided by Kuno
Kurzhals. (c) A static stimulus, like a powerwall display [441], inspected from a dynamic
position, allowing movement to change the perspective on the static stimulus. Image provided
by Christoph Müller. (d) A dynamic stimulus, like driving a car with many other cars and
pedestrians crossing our way while dynamically changing our positions [44].
4.3.3 Tasks
The tasks in a user study can come in various forms, ranging from simple
ones for which just one static diagram has to be inspected to very complex
ones demanding for a variety of interaction techniques including gestures,
touch, or gaze [413], applied to several visualizations, changing parameters,
letting run several kinds of algorithms, and understanding the components,
their interplay, and how they affect the dynamic user interface like buttons,
sliders, menus, and the like. Simple tasks are mostly required in controlled
studies in a lab for a technique-driven setting to evaluate if one static
visualization is better than another one [389]. The complex tasks, however,
are more interesting in visual analytics systems in which complex relations,
152 User Evaluation
(a) (b)
Figure 4.6 “Why is the road wet?” is a task that can be solved by watching a given abstract
visual depiction of a scene (a). The visual scanning strategy to solve this task has to follow
a certain visit order to grasp the information subsequently to solve the task (b). Eye tracking
can give some insights into such viewing behavior [416].
Figure 4.7 Varying the independent variable “link length” can have an impact on the
dependent variables error rate and response time for the task of finding a route from a start to
a destination node in a node-link diagram with a tapered edge representation style [97].
on them nor measure their impact on the dependent variables, we need some
kind of extra means to show the change to the human user, i.e. the visual
stimuli which serve as visual output. Those reflect the modifications of the
independent variables and get perceived, observed, and visually processed by
the user in order to efficiently solve a given task (see Figure 4.7 for a change
of the independent variable “link length” in a partial link study [97]). The
user efficiency comes as a challenging issue and it is impossible to measure
whether a user performed as efficiently as possible. However, to achieve a
real value for the performance measure we should trust the participants and
tell them to either respond as fast as possible or as accurately as possible.
Reaching both goals at the same time are typically two conflicting situations,
i.e. the faster we respond the more errors we make and the more accurate we
are the slower we are with the response. Most study setups ask for responding
as fast as possible while still keeping a high degree of accuracy, in cases
in which we are interested in the response time as a dependent variable
reflecting a performance measure to compare two or more visual variants
for example. In a comparative user study, for example, it typically does not
matter if the participants behave similarly for both settings, hence replication,
permutation, and randomization of the task blocks and trials should average
out this effect.
In a visual analytics system the independent variables can at least be
based on properties like the data, visualizations, interactions, algorithms, or
displays.
• Data based. Modifying properties of the underlying data in a visual
analytics system can show the impact on the dependent variables. For
example, given the fact that the data structure remains unchanged –
only its size in terms of the number of elements or the granularity gets
changed – might have an impact on a certain well-defined user task.
Another data aspect to be tested could be the completeness of the data in
terms of missing values and errors it contains. An independent variable
4.3 Study Design and Ingredients 155
could vary the extent of such data gaps, maybe with artificially generated
data based on a stochastic model.
• Visualization based. There are various visual variables that can be
varied to test their impact on user performance. For example, data
elements can be visually encoded differently focusing on lengths,
sizes, areas, shapes, colors, and many more. Moreover, layouts and
arrangements, visual complexities, compactness, or sparsity/density
properties can be adjusted to understand how they affect the user
performance. Even the positions of several views in a graphical user
interface can be modified which seems to be a meaningful independent
variable for a visual analytics system to be tested.
• Interaction based. If user interactions are allowed those can come
in a variety of ways [544]. Understanding the difference between two
or more interactions that focus on the same effect in a visualization
can help to pick the most effective and efficient one based on the
user performance. Moreover, input devices like a mouse, gesture,
gaze, voice, keyboard, or many more might be compared, or even
a combination of them to find the best way to interact in terms of
user performance. However, testing different input devices typically
demands completely different study designs and setups, hence the
results must be taken with care due to the fact that the impact might
not be caused by the input device but rather by the different study
setup.
• Algorithm based. A visual analytics system provides lots of algorithms
to process, transform, aggregate, project, or modify the given data.
Supporting several algorithms that actually produce similar results, or
one algorithm that produces several different results based on the same
input parameters, like a stochastic approach based on a random function,
could serve as an independent variable. It may be noted that the runtime
performance of the algorithm is not investigated here, but rather its
visual output that can be perceived and explored by the user based on
a certain task.
• Display based. The output device can also serve as an independent
variable. For example, small-, medium-, or large-scale displays provide
more or less space for the visual output, hence they might have an effect
on the user performance. Typically, a certain repertoire of interaction
techniques is required depending on the displays like mobile phones,
standard computer monitors, or large-scale high-resolution powerwall
156 User Evaluation
4.3.5 Experimenter
In a controlled study setting the experimenter guides the participants and
hence, plays a bigger role than in an uncontrolled setting in which the study
participants more or less have to guide themselves through the study trials.
The degree of control is higher if the experimenter is present and sitting next
to the participant, eager to help, like in a laboratory experiment. However,
the experimenter has to behave as constantly, smoothly, and similarly as
possible for each participant to avoid biases. Moreover, the experimenter
can be a confounding variable for several reasons, for example, just the fact
that a person has a different effect on another person makes the mimicking
of a similar behavior a difficult, nearly impossible challenge. Some study
setups even try to hide the experimenter from the participants, for example,
by a glass pane that is transparent on one side, but still the voice of the
experimenter can have an impact on the performance of the study participants.
Furthermore, even if the experimenter tries very hard to behave the same for
every study participant, it is just not possible due to mood swings and personal
reactions to certain participants. Experimenters are human beings and hence
prone to feelings, errors, and misunderstandings. The only way to mitigate
this situation is to average out the effects in the performance measures by
testing many participants while, hopefully, the experimenter does not have a
positive or negative impact on all of them in the same way.
Also in an uncontrolled study the experimenter has an impact, for
example when interviewing people later on. However, in running an
uncontrolled experiment, the study participant is typically left alone, for
example in a crowdsourcing experiment, and the experimenter has to rely on
the participants’ best performances without giving them further instructions
or giving them the chance to ask for help. The experimenter has neither the
opportunity to intervene, in case the participants do not follow the study plan,
nor does the experimenter have a good way of observing the participants or to
take notes, unless the session is recorded as audio or video. In an uncontrolled
setting it should be avoided that the experimenter is seen. Normally, it is
not needed because all the instructions can be given in written form that
have to be read by the study participants. However, also the writing and
explanation style might have an impact on the participants, but at least it is
the same style; however, the interpretation is based on the personal attitudes
and understandings of the study participants. In summary, the experimenter
plays a role to different extents when taking the perspective of before, during,
or after a running experiment.
158 User Evaluation
Figure 4.8 Since eye movement data is composed of at least three data dimensions like
space, time, and study participants, the visual representations also get more complex with
many aligned and linked visual components supporting pattern identification in the data [298].
Here we see the x–y positions in the top row, the saccade lengths and orientations in the center
row, and the filtered pairwise fixation distances in the bottom row while time is pointing from
left to right.
160 User Evaluation
about the major distribution of the numbers. The mean value, i.e. the average
value, takes into account all numbers, however, it can be a value that is far
from being a representative value of the distribution, for example in a case, in
which the numbers are distributed at both ends of a scale and not in between.
On the other hand, the median value, which could be explained as the value
in the middle, could also not tell anything about the distribution of the list
of values. However, it is nice to have such descriptive statistical values that
give an impression about a central tendency of a value list. They already give
some very general hints about the data, for example, when inspecting the
minimum and maximum in combination which provides insights about the
range of the values. Negatively, they do not tell us enough about a distribution
of the values and can lead to wrong conclusions, like in the example of the
Anscombe’s quartet.
To provide even more insights into a list of user study performances,
statistics is equipped with further expressive values, for example, indicating
the distribution or spread of a list of values around a certain point or in a
certain value range, but telling us more than just the standard minimum,
maximum, mean, or median values. The variance, for example, takes into
account each value by including the difference to the mean value which is
important for the spread of the values in a list. To reach this goal, the sum of
the squared differences of all values to the mean divided by the size of the
value list gives the variance (Var) which is at the same time also the squared
standard deviation (SD), a more commonly used term in a scientific report or
researchPnpaper. If X := {x1 , . . . , xn } expresses a list of performance values,
xi
x := i=1 n the mean value, then the variance and standard deviation are
defined as
Pn
2 (xi − x)2
Var(X) := σ = i=1
p n
SD(X) := σ = V ar(X).
Generally, the variance describes the spread effect of the values in a given
list from their mean value. It may be noted that if we use a population for our
performance value list, we use n to divide, while for a sample, i.e. a smaller
part of a population, n − 1 is used.
A histogram (see Figure 4.9(c)), on the other hand, can be used to show the
distribution of a population with respect to the performance measure, i.e. a
quantity is mapped to the x-axis instead of a categorical information, the
participants, as in the case of a bar chart. In a histogram we typically encode
the number of people falling into a certain bin, i.e. a value range on the x-
axis representing the corresponding performance values. The height visually
depicts this number and hence, gives an easy-to-understand overview of the
bins that are frequently hit and those in which not many values fall. The
shape of the resulting histograms can be interpreted for patterns indicating
a property of a certain distribution (see Figure 4.10).
If the evolution over time of a performance value is of interest we might
use line charts that connect the points indicating a certain value at a time
point (see Figure 4.11). This mostly results in visual shapes that help to
identify trends or countertrends, in case several of those temporal variables
are plotted, for example, focusing on fatigue or learning effects that might
have an impact on a performance measure over time. Moreover, temporal
data might include anomalies or outliers, i.e. effects that do not follow the
overall trend pattern. It may be noted that line charts should be taken with
care if discrete or categorical values are depicted at the x-axis since the lines
would reflect some kind of interpolation effect that might let us perceive
non-existing values between the discrete time steps or even between two
neighbored categories. However, the shape created by adding lines to the
points is perceived as some kind of closed curve and hence comparisons to
other such curves are perceptually easier than if only the point set would be
visible. This is due to the strengths of the Gestalt laws of good continuation
and closure [292].
Depending on whether univariate, bivariate, trivariate, or multivariate
data is measured, we have to rely on different types of visualizations to
show patterns in the data. For example, for univariate data, i.e. data that
is composed of just one variable, like the response time, we could show a
4.4 Statistical Evaluation and Visual Results 165
Figure 4.11 A line chart is useful to depict several time-varying performances to identify
trends as well as countertrends and to compare them for differences over time.
Figure 4.12 A box plot can show the distribution of a univariate dataset, for example the
performance measure of the response time or the error rates.
histogram, or a box plot (see Figure 4.12). The box plot is useful to show the
general spread of the univariate data on a quantitative scale. It indicates data
values like the median, which is the middle value of the dataset, the minimum,
the maximum, the first and third quartiles which are the medians of the lower
and upper halves of the dataset. Moreover, if correlations of two variables are
of particular interest, for example, between an independent and a dependent
variable or between two dependent ones like the response time and the error
rate, we typically show this bivariate data in a scatter plot (see Figure 4.13(a)).
Correlations between more than two variables, for example tri- or multivariate
data, are depicted as scatter plot matrices (SPLOMs) (see Figure 4.13(b)) or
parallel coordinates plots (see Figure 4.13(c)). The most prominent patterns
to be visually derivable are positive or negative correlations, for example, a
longer response time might cause more (positive correlation) or less (negative
correlation) errors.
166 User Evaluation
(a) (b)
Figure 4.14 A scatter plot enriched by error bars indicating the standard error of the means
(SEM). The average saccade length is plotted on the y-axis while the average fixation duration
is shown on the x-axis. (a) The complexity levels. (b) The task difficulty.
Figure 4.15 The Euclidian distance to the start is plotted over time to show the progress of
visual attention with respect to such a relevant point of interest in a visual stimulus.
study is based. Moreover, the output in terms of statistical results, how they
are computed, and how those results are visually depicted will be discussed.
It is worth noting that in this section we only focus on user studies without
the explicit use of eye tracking as a technology to record visual attention
which would go beyond the scope of this section, but those will definitely be
discussed later in Section 5.4.
the assumption that the provided tree visualizations complement each other
since no individual one can focus on all aspects in the hierarchical data such
as depth, size, or branchings [494]. Moreover, it was partially investigated
whether several of the techniques in combination can be more powerful
than one technique alone. A year before, only three tree visualizations,
RINGS, treemaps, and Windows Explorer were compared [520] by asking
18 participants. Qualitative ratings as well as task completion times were
measured. Another comparative study using the hierarchical visualization
testing environment [13] also looked into four hierarchy visualizations
by recruiting 32 participants answering eight tasks. The visualizations
under investigation were Windows Explorer, the information pyramid, the
treemap, and the hyperbolic browser while the response times and subjective
ratings were recorded. In a similar direction we find a study focusing
on six tree visualizations [291] with 48 participants. This study uses
Windows Explorer as a baseline comparison system and quantitative as
well as qualitative measures were recorded. Only three tree visualization
techniques were compared for hierarchies with large fan-outs [471] by
recruiting 18 participants. Response times, error rates, and verbal feedback
was recorded for various hierarchy-related tasks. There are many more
hierarchy visualization user studies, focusing on different independent
variables checked for a variety of tasks. Such studies focus on hierarchies
in source code [21], treemaps vs. wrapped bars [536], 2.5D treemaps [333],
progressive treemaps [425], as well as combined treemaps [331], node-link
trees [398], a space-reclaiming variant of the icicle plots [509], indented
trees [192], or H-tree layouts [436], to mention a few.
The visual metaphor for graphs can be investigated, for example, asking if
node-link diagrams or adjacency matrices are better in terms of performance
measures. One study asks for typical tasks [200] focusing on understanding
which one of the aforementioned approaches produces better results in a
comparative study. Thirty-six participants were recruited while the error
rates and response times were measured as performance indicator. Another
comparative study [238] tries to understand which kind of edge representation
styles like animated, tapered, curved, or standard with arrow head (see
Figure 4.17 for a few examples of edge representation styles) is suitable for
the task of finding connected nodes in a network. Twenty-seven participants
answered various trials while their response times and error rates were
recorded. Also qualitative feedback was investigated to understand people’s
preferences. Moreover, partial links [97] were researched to understand
the shortest link length that still gives good performance results. Forty-
two participants answered path finding tasks in a controlled laboratory
study while error rates and response times were recorded. The crossing
angles effect was checked by 22 participants in an uncontrolled online
study setting [248]. The response times were recorded and those of the
correct answers were analyzed. Brief interviews provide some qualitative
feedback. Finally, the dynamics of graphs with respect to memorability tasks
was investigated [17]. Error rates, response times, and the 25 participants’
ratings were measured and recorded. There are many more studies related to
graph visualization, some even with eye tracking, which is less prominent
in hierarchy visualization, in particular focusing on aesthetic criteria [29,
407] like the impact of link crossings [247], the layout [400], or edge
representation styles such as curvature effects on link interpretation [136,
534]. Moreover, some studies focus on connectivity models with respect
to node-link diagrams and adjacency matrices [6, 280], while others take
into account the mental map in dynamic graph visualizations [16, 408],
4.5 Example User Studies Without Eye Tracking 171
(a) (b)
Figure 4.18 Using interaction techniques to adapt parameters in a contour line-based visual
attention map. The public transport maps of (a) Zurich, Switzerland and (b) Tokyo, Japan.
between two different study settings, i.e. with and without interaction in a
comparative study setting. Moreover, other tasks focused on understanding
if the interactions are useful at all and can be applied by the participants in
a suitable way. Response times and qualitative feedback were recorded and
evaluated. In another study focusing on interactions for exploring dynamic
graphs [69], 20 participants were recruited to examine if the interaction
techniques incorporated in a dynamic graph visualization tool are intuitive
for the study participants to solve a given task. This task was too complex
to be solved by the static version of the tool without interactions. Response
times were recorded and qualitative feedback was requested. In the same
line of research focusing on dynamic graph visualization we can read about
the evaluation of two interaction techniques [181]. Sixty-four participants
answered tasks based on interaction in a controlled study while error
rates and response times as well as qualitative feedback was recorded.
Also eye movement data visualization is in the focus of user evaluation,
in particular the combination of several views [100] and interactions for
modifying the individual views as well as linking them together. An
uncontrolled experiment showed the usefulness of the interactivity of such
an eye movement data visualization tool while the qualitative feedback of
10 participants was analyzed for insights on the usability of the tool and its
functionality with respect to find patterns and strategies in the visual attention
behavior of people when inspecting public transport maps [372]. Interactive
timeslicing, animation, and small multiples were investigated in several
user studies for visualizing dynamic graphs on large displays [322]. For all
study setups the same 24 participants carried out tasks while their response
times, error rates, and qualitative feedback were measured. There are many
more studies with a clear focus on interaction techniques in a visualization-
or visual analytics-related context. Those range from preferential choices
based on a set of interactive visualization techniques [28], Fitt’s law with
respect to the design of interactive user interfaces [350] based on pointing
devices, or navigation tasks for off-screen targets on mobile phones for which
the display space limitations generate challenges for standard visualization
techniques [27].
Eye tracking has become a well-studied field these days [161, 235], not only
because of the progress of the hardware technology and the reduced costs for
eye tracking devices, but also due to the various application fields that benefit
from it. Such fields are numerous with varying backgrounds like marketing
that is interested in the user behavior to improve the selling strategies,
visualization and visual analytics which are focusing on understanding the
interplay of their algorithmic, visual, and interactive components to identify
design flaws and to build a starting point for enhancements, or software
engineering trying to figure out how software developers produce source
code and how efficiently and effectively they implement or debug such
code while they interact with a software development environment or while
they collaborate and communicate with other software developers. Moreover,
fields like neuroscience, cognition, human–computer interaction, medicine,
sports, or the automobile industry all benefit from eye tracking, given the fact
that the data is measured accurately and analyzed afterwards.
To efficiently and effectively record eye tracking data, a profound
knowledge of the anatomy of the eye is needed. Moreover, the internal
processes and concepts involving facts about light, vision, perception,
cognition, and psychology have to be researched and understood in order to
obtain reliable and significant results based on eye tracking data. The human
eye is very complex but to build a simple and cheap eye tracking device that
is powerful enough to measure and record eye movements, even on a very
coarse-grained and not very accurate level, at least the major components
of the eye have to be studied as well as their connections, impacts, and
causes for certain effects. Such effects also include diseases that might have
a negative influence on the way we measure the eye movements but also on
the trustworthiness of the results. Hence, the eye is involved in the field of
eye tracking in many respects, it is not just “the window to our soul”, it builds
some kind of interface between the visual stimuli and the cognitive processes
happening in the brain.
175
176 Eye Tracking
The field of eye tracking has existed for many years but the technology
as we know it today is not comparable with the early attempts to observe
people’s eyes while they were solving a certain task. Although there was
a general impression about what eye movement looked like, it was pretty
challenging to note down the various numbers of movements, i.e. just very
general observations could be made. The invention of the computer and the
progress in hardware technology in general has led to the eye tracking devices
that we know today with the ability to record eye movements very accurately
as well as additional physiological measures. Moreover, the steady progress
in software technologies, in particular data analysis, visualization, and visual
analytics, give great support in helping us to identify patterns in the eye
tracking data, even in real-time. Compared to the early attempts we can say
that we are eye witnesses to the great advancements in the field; however,
other, even more challenges, occur these days, moving away from the data
recording to more data analysis challenges. Describing the many concepts
that exist for analyzing eye tracking data is beyond the scope of this chapter,
hence we move that to another part of this book which is located in Chapter 6.
Moreover, interacting with the eye, instead of using devices like the
mouse, joystick, keyboard, or further concepts based on touch, gestures,
voice, and the like, brings into play another kind of giving user input to an
interactive visual stimulus. Such gaze-assisted interaction is researched as a
novel discipline but although it seems to be promising, it also brings new
challenges into play, for example the well-known Midas touch problem that
describes the effect of making everything that is visible in a visual stimulus
interactable by focusing on it with the eye. However, this effect is counter-
productive since it generates some kind of over-reaction in the users’ visual
attention, hence good ways out of this dilemma have to be developed. This
issue requires special gaze-assisted features which are not known in the same
way for other interaction devices and methods, but if several interaction
concepts are combined in a clever way, they seem to create a possible solution
for this problem. However, user evaluation is required that brings another kind
of data to be analyzed into the field.
In particular, applying eye tracking to visual analytics systems requires
the knowledge of many concepts due to the interdisciplinary character of such
systems ranging over, and including, general fields like algorithms and data
structures, human–computer and gaze-assisted interaction, visualization, as
well as cognition, perception, or psychology, to mention the most important
ones. The interdisciplinary fields are one reason for the generated eye tracking
dataset coming with a certain degree of complexity. No matter how complex
5.1 The Eye 177
the data is, it is worth analyzing and visualizing since it hides interesting
and valuable patterns, correlations, strategies, rules, or associations that
correspond to a certain user behavior. The understanding of such visual task
solution strategies depends on how well we are able to translate the abstract
eye tracking datasets into a language that easily explains the human behavior
while solving a task. These insights can finally help to identify negative
issues like design flaws to hopefully enhance a static, dynamic, interactive,
animated, 2D, or 3D stimulus based on such advanced analyses.
Figure 5.1 The human eye is a complex organ that is important for the visual system [196].
Moreover, it builds the major ingredient for all eye tracking studies.
rods, cones, and optic nerve (see Figure 5.1). These play crucial roles when
setting up and conducting eye tracking studies. Actually, the human eyes have
many similarities to digital cameras. The cornea which takes the role of the
camera lens is hit by light. The iris imitates a diaphragm, similar to that of a
camera which is responsible for increasing or decreasing the amount of light
that falls into the eye, in particular, the light ending up at the backside of
the eye, i.e. the retina. To reach this goal the pupil’s size can be modified
in order to regulate the amount of light falling in. The eye’s lens plays the
role of focusing visual objects just like the autofocus mechanism of a digital
camera lens. The remaining light reaching the retina which is consisting of
rods and cones, transforms this light into electronic signals, transmitted by
the optic nerve to the visual cortex which stands for the region of the brain
that manages the sense of sight [521]. From here, cognitive processes guide
the actions and reactions of the human body functions, also the movement of
the eyes.
The rods and cones placed on the retina take the role of photoreceptors.
Several million of them [196] absorb light to transform it into nerve impulses
which are further sent via the optic nerve to the corresponding brain region.
One difference between rods and cones is their effectivity at different day
and night times. Rods are important for vision at night since they are
more sensitive to light than cones, hence if the light is very low the rods
play the most crucial role for human vision. The cones instead, are useful
during daytime when the light is very bright, i.e. there are many more
photons than during nighttime. Rods and cones have an impact on color
perception, but this impact is highest for the cones. Certain wavelengths,
typically short, medium, and long ones, are considered, characterizing the
cone photoreceptors into three classes. Research about these aspects describes
the number of photoreceptors in the human retina as being approximately
6 million for the cones and 120 million for the rods, but these numbers
vary from eye to eye. For eye tracking studies, it is important to use
bright light in case the cones are more the focus of the results whereas
in a comparative eye tracking study investigating how different light levels
affect the eye movement behavior and further dependent variables we need
profound knowledge about the functions of the photoreceptors [253].
(a) (b)
Figure 5.2 Cataracts [392] affect the lens of the eye in some kind of degeneration process
causing clouded and unclear vision: (a) clear vision; (b) an eye with cataract issues.
study. In particular, for long-duration tasks, which might also occur in visual
analytics eye tracking studies, this effect can have negative consequences on
the reliability of the results. Finally, with eye tracking we might find out the
emotional states of people [491].
For standard eye tracking studies it would be too expensive or even
too frightening for people to take part if they have to visit a doctor or eye
specialist first to check for diseases, hence the number of participants would
be reduced tremendously for studies investigating issues in visual analytics.
The best way is, consequently, to just invite the people for the study and ask
them if they would like to fill out a corresponding form while explaining them
the ethics and privacy issues related to this information providing process. In
any case, it is good advice to check participants beforehand, i.e. they should
fill out some kind of form that includes information about personal details
to which diseases also belong. However, such private information should not
be abused and hence it should be explained to the participants if they decide
to provide such information. Moreover, the recorded personal data should be
anonymized in a way that it cannot be recovered later on.
(a) (b)
(c) (d)
Figure 5.4 Refractive errors: (a) nearsightedness (myopia) and (c) farsightedness
(hyperopia) can be corrected by special lenses (b), (d).
conclusion from such experiments was that, depending on the task, the
spectators perform different scanpaths, i.e. their eye movements can vary a
lot. However, although the recording of the eye movement data was quite
accurate in these days, the quality of the data cannot be compared with
that generated by eye tracking devices having undergone many stages of
technological progress as we have them today. The devices to record the
eye movements in these old days were mostly based on so-called suction
caps [492], similar to contact lenses as we know them today for correcting
refractive errors in the eyes.
The progress in the hardware technology brought various novelties into
the field. With those novel inventions and more and more accurate eye
tracking devices, many more challenges occurred, mostly focusing on the
analysis concepts needed for a proper and insightful evaluation of the
recorded eye movement data combined with additional data sources like
physiological and verbal data [44]. Moreover, visual analytics was detected
as a field to further uncover data patterns by including the human user
with interactive tools containing visual components as well as algorithmic
ones. But still, the field is moving forward since many open challenges still
remain such as the link between visual attention, cognitive processes, and
psychological issues that might guide the visual task solution strategies [305].
With steady progress we have reached a level at which even real-time eye
movement data can be recorded and analyzed. Moreover, eye tracking can be
used as an interaction modality, normally known as gaze-assisted interaction,
demanding for very accurate eye trackers, depending on the interactive visual
components provided by a user interface, for example, a complex visual
analytics system.
Figure 5.5 Eye movements during a reading task consist of short stops (fixations) and rapid
eye movements (saccades). This insight was found by Hering, Lamare, and Javal around
1879 [263].
later replaced by photographic tape. Also in their method it was only possible
to record horizontal eye movements. In 1905, Judd, McAlister, and Steel were
the first who presented a device that was able to record eye movements in both
directions, i.e. horizontal as well as vertical with the drawback that the study
participants had to sit or stand still during the experimentation. The most
popular outcomes of these days, although eye tracking was in its infant ages,
were that humans do not obtain information during rapid eye movements
(saccades), that the eyes need some time to first initiate before they can start
obtaining information, and the fact that humans have a limited visual field that
only allows focusing on the visual objects in the center of the view sharply
while the objects at the edge of the visual field are more observed in a blurred
fashion [521].
Figure 5.6 An example of an eye tracking device as we know it today, known as the Tobii
Pro Glasses 3. Image provided by Lina Perdius (Tobii AB).
typically touch or have a contact with either the eye or the skin around
the eye. This contact makes such studies uncomfortable for the study
participants and in rare cases, at least in early times, people might have
suffered from some eye injuries or at least their eyes had to relax some
time after having taken part in such an eye tracking experiment. On the
other hand, systems that rely on technical equipment that actually avoid
these contacts try to make the study procedure more comfortable for the
subjects, at the cost of finding a good way to get accurate and reliable eye
tracking measurements, for example, if head movements in interactive
applications must be allowed [360].
• Remote vs. mobile/head-mounted/wearable/portable. Looking at
where the eye tracker is placed and how much flexibility it provides
brings us to another categorization consisting of major classes. Those
contain devices that are used in remote settings, placed away from a
study participant without direct eye or body contact allowing contactless
measurements, typically integrated into a computer monitor. The eye
tracking devices could even be mobile eye trackers in the sense of being
applicable everywhere, for example, in field studies, in which a high
degree of flexibility is required. Moreover, such systems can be head-
mounted, wearable, or portable which is achieved by better and better
hardware technologies. Some mobile systems are so advanced that they
can even track the eye movements in combination with head movements
to reliably compute the points of eye fixations.
• Electro-, infrared-, video-oculography, and scleral search coil.
Another classification is based on how the eye movement data is
acquired. For example, certain sensors might be placed around the eye,
measuring the skin potential during eye movements based on electric
fields called electro-oculography. Also infrared light can be thrown on
the eye to better record the effects of eye movements on the pupil
positions. The amount of reflected light and, in particular the amount
changes play a key role in this approach called infrared-oculography.
Video-oculography, on the other hand, takes into account the video-
recorded images based on single- or multiple-camera eye trackers. The
corneal reflection technique makes it possible to record the position of
the pupil given by the additionally reflected light. Finally, the search coil
method uses wires in some kind of contact lens placed in a magnetic field
causing voltages based on Faraday’s law which give a hint about the eye
position.
192 Eye Tracking
Table 5.1 Eye tracking companies with respect to hardware and software developments as
well as focused applications, described by major buzz words
Company Developments/Technologies Applications/Focus
Argus Science Binocular, 3D mobile, real-time Marketing, sports
Blickshift Visual analytics, data analytics Usability studies
Ergoneers D-Lab, Dikablis glasses, portable Automotive, marketing
EyeSee Online, facial coding, webcam Virtual shopping
EyeTech USB-connected, low-powered Assistive, disabled
EyeVido Browser data, data analytics UX, usability studies
EyeWare 3D software, depth camera Real-world interaction
GazeHawk Webcam, crowd sourcing Usability, comparisons
Gaze Intelligence MRI, mobile, remote Behavioral studies
Gazepoint Biometrics, GP3 HD UX design, usability
iMotions Biometric sensors, real-time Human behavior, UX
ISCAN Real-time, head-mounted Pilots, military
LC Technologies Eyegaze, tablet communication Assistive, disabled
Mirametrix USB, attention sensing UX, HCI
Pupil Labs Open-source, wearable headset UX design, marketing
Smart Eye Head tracking, AI-powered Automotive, aviation
SMI Glasses, VR, RED500 Research, neuroscience
SR Research EyeLink 2, portable Academic research
The Eye Tribe Tracker Pro, smart phone Gaming, web usability
Tobii VR headsets, wearable Usability, VR, cars
to sell a product. Eye movements can at least give a hint about the visual
attention strategies; however, they do not tell us what the customers
are thinking, i.e. cognitively processing. There is a lot of research in
this domain based on eye tracking with the goal of understanding the
customers viewing strategy and, based on that, improve a product design
or even the placement of products in a department store or on a web
page [23].
• Immersive analytics/VR/AR/MR. Virtual, augmented, and mixed
reality environments stand for some novel technologies in the field of
data analytics, for example in the field of immersive analytics [347].
Combining them with eye tracking technologies can lead to powerful
tools in the domain of data analysis while the human users are even
more integrated in the data analysis process by taking into account
their eye movements [140], even as a way to interact with the system
by using the eyes as in gaze-assisted interaction. For example, recent
technologies like the HTC Vive Pro Eye or the Microsoft Hololens
2 integrate eye tracking to improve the user experience. However, to
allow a combination of eye tracking data with the rest of the system,
real-time eye movement data analysis is required, supported by efficient
algorithms running in the background.
• Gaming/entertainment/sports. The fields of gaming and entertainment
benefit from gaze-assisted interaction in a way that the user has an
additional means to interact, which makes the gaming experience
much more realistic, for example if target-based shooting or foveated
rendering is based on the gaze input. However, on the negative side, it
is also more difficult to learn how to eye-control the games, even how
to combine different interaction modalities. Moreover, analyzing and
visualizing the recorded eye movement data [83], for example to explore
visual attention strategies of the players for patterns and anomalies, is
a challenging task due to the time-varying, interactive visual stimulus,
even in a collaborative way as in multiple player systems. In the field of
sports we might use eye tracking technologies to examine how active
players perceive their environments during a match, for example to
improve their winning strategy based on the detection of inefficient
eye movements or the fact that they have been unaware of a certain
situation [252].
• Education. We could inspect the teaching process from two
perspectives, from the teacher and the students’ side, for example who
pays attention to what [147]. Such eye movement data could give
196 Eye Tracking
Figure 5.7 Eye tracking technologies can be useful in the field of aviation, in particular,
when training pilots to land a plane [430, 432]. Image provided by David Rudi (Copyright
ETH Zurich).
Figure 5.8 Eye movement data can be described as consisting of gaze points, which is the
lowest level of granularity that is interesting for eye tracking in visual analytics. Those gaze
points are spatially and temporally aggregated into fixations by modifiable value thresholds.
The fixations with duration (encoded in the circle radius) contain saccades in-between, i.e.
rapid eye movements. A scanpath is made from a sequence of fixations and saccades. Regions
in a stimulus that are of particular interest are called areas of interest (AOIs). If we are only
interested in fixations in a certain AOI we denote those by gazes. Between AOIs there can be
a number of transitions indicated by the number of saccades between those AOIs [44].
The stimuli are a key ingredient in an eye tracking study since they carry
semantic information that has to be understood and combined in a way to
successfully solve the given tasks. These stimuli can have several properties
and make a difference for the data storage, linking, as well as later data
analytics in the form of algorithmic as well as visual concepts combined
with interaction techniques, such as visual analytics. The stimuli can be just
static or even dynamic, i.e. changing over time as in a video. Moreover, if
the content is changing on user demand, like in interactive user interfaces,
we can have many different static stimulus states that might form some kind
of stimulus graph [68] instead of a linear sequence of animated frames or
snapshots from a video in which each frame is typically watched only for
a fraction of a second. In general, one challenge will be the linking of the
stimulus data with the eye movement patterns before starting an analysis.
Moreover, combining several different stimuli with similar characteristics
attached by eye movement data is a challenging synchronization task as well
as the storage of the data if it comes to long-duration eye tracking tasks in
the wild where video sequences have to be recorded together with the eye
movement data.
5.3 Eye Tracking Data Properties 199
In this section we will have a look into the different stimuli in eye tracking
studies. As a next step we describe fixations and saccades that model the
visual scanning behavior of several people, for example when taking part in
an eye tracking study. Areas of interest (AOIs) are useful to explore the visual
attention patterns based on aggregated regions in a stimulus. Static AOIs
are easier to handle than dynamic ones, for which some kind of matching
function is required. Apart from the pure eye tracking data we can easily
enrich the data by extra physiological measures such as EEG, galvanic skin
response, or pupil dilations, to mention a few. Finally, additional metrics
can be derived from the original data to get more insights from a different
perspective, typically on a more aggregated view or in a form that links
different data aspects to build a new kind of measure, for example the ratio
of fixation durations and saccade lengths, or the time to first fixation to areas
of interest. Actually, there is no limitation to what types of data can be used
as extra input, it is just a question of what is meaningful for solving the tasks
at hand from a data analysis perspective. Moreover, the runtime complexities
for certain algorithms might need to transform the data in the desired format
first which could have an impact on a possible interaction in a visual analytics
system for analyzing eye tracking data.
eye movements are interesting since they describe some kind of path that is
taken in the visual stimulus which hints at facts like how we solved the task,
where we stopped for longer and shorter times, and if certain visual objects
have been inspected several times in a row. However, the eye movements do
not describe what is cognitively processed in the brain, they just give hints
about the visual attention patterns.
Visual stimuli in an eye tracking study can come in several forms.
Figure 5.9 shows an example of a dynamic stimulus from a car driving
experiment.
• Static visual stimulus. A visual stimulus could be a static diagram,
picture, poster, or any visual object that is not equipped with interactions
and that cannot be rotated or for which there is no opportunity to walk
around to inspect it from different perspectives. Typical examples are
standard “old” diagrams before the invention of the computer or those
for which interaction is turned off (see Figure 5.8) like in a technique-
driven user experiment. Moreover, text could fall into this category for
which researchers try to investigate reading tasks and the corresponding
eye movements, also static scenes like pictures or paintings for which
a task is asked, like in the work by Yarbus [539]. To keep the setup of
an eye tracking study easy, as well as the data analysis later on, in most
cases a remote eye tracking system is used, for example integrating the
eye tracker in a computer monitor on which the static visual stimulus is
shown.
• Dynamic visual stimulus. A dynamic stimulus is one that changes its
content from time to time. This does not mean that the stimulus itself
has to change necessarily but it could be possible that an observer has
the possibility to watch the stimulus from many different perspectives,
for example, walking around in a museum and inspecting a “static”
sculpture. This visual object is always the same but the viewer’s field of
view changes dynamically while walking closer and farther away, also
5.3 Eye Tracking Data Properties 201
with varying gaze depth [163, 166]. In a shopping task we have a similar
situation, but here the viewer typically explores a larger dynamic scene
while other customers cross the way and we might start conversations
and communications. In a car driving task we see the dynamic stimulus
passing by while at the same time navigating the car and also interacting
with several electronic devices like the radio, telephone, or navigation
system. Similar aspects hold for virtual environments in which people
freely walk around or react to certain situations with body movements or
gestures. Moreover, a video or animation belongs to the class of dynamic
stimuli although the viewers cannot change their content, maybe just
pause and play back. It is more like an autonomously changing linear
sequence of static pictures. In visual analytics or many other application
fields we come across interactive user interfaces which are by definition
dynamic. For such a dynamic visual stimulus we require more advanced
head-mounted, portable, or wearable eye tracking devices. However,
analyzing such data is much more challenging because the dynamics
of the stimulus has to be stored by a video and this demands for
algorithmic concepts like dynamic AOI detection, matching and linking
of the stimuli over time, or just identifying the group behavior and visual
attention paid over space and time.
A visual stimulus can be mathematically modeled as a sequence of static
pictures carrying a time point given as a natural-numbered index, no matter if
the stimulus is static or dynamic. In a static case the sequence consists of the
same stimulus all the time. This means the visual stimulus can be modeled as
S := (S1 , . . . , Sn )
K := (K1 , . . . , Kn )
for n ∈ N describing again the total amount of time the stimulus was visually
attended. Those knowledge states might be updated from time to time, but
we cannot really describe how they are updated; however, they might have an
impact on the visual attention strategy [269].
of all gaze points in a gaze point cluster. There are even more advanced
methods [463] which are not required for eye tracking visual analytics. The
fixation radius is typically so small that it is irrelevant where exactly inside
the gaze point cluster the representative fixation point is located.
The total amount of time a fixation lasts is called fixation duration
which can be interpreted as the longer a visual object is fixated, the more
interesting it is and it attracts our attention; but on the other hand, it might
also be confusing and some time is wasted in understanding its meaning.
The eye movement from one fixation to the next one is called a saccade
while the sequence of all fixations and saccades forms a scanpath (see
Figure 5.8). However, there are also smooth pursuits that describe eye
movements constantly following a moving visual object, i.e. saccades are not
occurring very often due to the fact that all gaze points are close together
over time and space by a small constantly shifting offset. In rare cases there
might occur so-called catch-up saccades which are effects caused by visual
objects that move too fast for keeping an eye on them or a distracting visual
object that also attracts the viewer’s attention at the same time. A scanpath
describes a trajectory over space and time making it a challenging type of
data to analyze if several people’s scanpaths are involved in finding a common
visual scanning strategy in the spatio-temporal data.
Mathematically, we can model a sequence of gaze points as
G := (g1 , . . . , gk )
F := (p1 , . . . , pm ),
we get much fewer points, i.e. m < k, depending on the given parameters.
Each pi , 1 ≤ i ≤ m describes a point in 2D as pi ∈ X × Y , similarly
to the gaze points. However, each pi is based on a group of gi and, hence,
some kind of aggregated or representative value based on gi . Moreover, each
pi is attached by two time points tei and tli , expressing the time points the
eye enters the fixation space and when it leaves it again. Consequently, each
fixation has a fixation duration given by tdi := tli − tei in addition to the
spatial position on the stimulus. Between each pair of subsequent fixations
204 Eye Tracking
we have a saccade that describes the eye rapidly moving from one fixation
to the next one in the sequence. All fixations in a fixation sequence and the
saccades in between together form a scanpath. If several people are taking
part in an eye tracking study we have many of those scanpaths, typically all
with varying properties incorporating the visual attention behavior of many
people. The sum of all fixation durations
m
X
tresponse := t di
i=1
(a) (b)
Figure 5.10 Selecting areas of interest in a static stimulus can reduce the amount of eye
movement data and can impact the eye movement data analysis since each AOI is some kind
of spatial aggregation: (a) AOI selection based on hot spots of the visual attention behavior;
(b) AOI selection based on the semantics in the stimulus [100].
can happen smoothly, but also abruptly, jumping from one position to a
completely different one. Moreover, the visual object to be located inside an
AOI can change its shape dynamically, it can rotate, get bigger or smaller, be
partially occluded for a certain amount of time, and so on. All of these effects
make an automatic detection of an AOI a difficult and sometimes error-prone
process. However, defining dynamic AOIs manually is possible, but it is a
very time-consuming process.
There are several ways to define areas of interest that are independent of
the fact whether a stimulus is static or dynamically changing over time (see
Figure 5.10 for an example of such an AOI definition based on user input).
• User input. Manually selecting areas of interest is a possible solution.
For static stimuli this sounds doable, but for dynamically changing
AOIs this can become a time-consuming process. However, a manual
user-specified AOI definition and selection can be more exact than an
automatic one due to the perceptual strengths of the human visual system
and pattern recognition abilities [521]. Moreover, the selection can be
based on rectangular, polygonal, or arbitrary shapes.
• Automatic spatial subdivision. A stimulus, static or dynamic, can be
split into several subregions based on equally sized grid cells while the
cell sizes are modifiable and adaptable by the user. Although this is a
naive, semantics- and visual attention-ignoring process, it works quickly
and might give some hints about the eye movement data. However, a
more fine-tuned AOI definition should follow after this naive idea is
applied.
206 Eye Tracking
of the input information has been covered by the visual attention, which can
also be estimated by the enclosed region of the fixation points in a scanpath,
i.e. given by a minimal polygonal shape that encloses all fixation points.
Also areas of interest provide a way to form some kind of derived metric
since they all contain additional insights for specific stimulus regions like
the numbers of fixations to certain areas which allow comparisons between
stimulus regions, even for comparing participant groups based on their visual
scanning behavior. Moreover, metrics like time to first fixation, time spent in
an area of interest, number of AOI revisits, number of transitions between
two AOIs, and many more provide interesting additional values.
If several metrics are combined we can further extend the repertoire of
possible metrics; however, since this repertoire is very large, we might better
consider which ones make sense for a certain data analysis task beforehand
and then select the ones that are interesting for these specific tasks. For
example, a combination often seen is the ratio between fixation duration
and saccade length, or the saccade orientation (in degrees) divided by the
saccade length, average saccade length in a certain area of interest, average
distance between all pairwise fixation points, and many more. All of these
derived metrics serve their own purpose and can be analyzed for further
statistical properties and also for correlations between two or more of them,
for example, asking oneself if the fixation duration and length of subsequent
saccades are positively or negatively correlated [299].
Figure 5.12 Public transport maps for different cities in the world (in this case Venice in
Italy) [372].
visual stimulus used to efficiently solve such route finding tasks. By using
eye tracking we might find out if the map contains certain design flaws [372],
if its design is better or worse than another one for the same city, if the map
complexity in terms of number of stations and metro lines plays a role for the
visual attention strategies [85], or if visual enhancements in the form of color
coding, legends, or sights [65] have any impact. Twenty-four public transport
maps from cities all over the world were used as stimuli by changing between
colored and gray-scale stimuli to investigate if color has any impact on the
user performances as well as visual task solution strategies [372]. Route
finding tasks were asked in this study by highlighting the start and destination
stations to reduce cognitive effort when finding those relevant points for
answering the task (see Figure 5.12 for a public transport map overlaid with a
scanpath). Forty participants were recruited and a Tobii T60 XL eye tracking
device was used. The major result showed that the original task of finding
a route between two highlighted stations was actually subdivided into many
subtasks [63] containing cross checking behavior, i.e. the found route was not
confirmed immediately but after a careful check of the found route. Moreover,
color coding was found to be very important for lowering the response times,
while the saccades became much shorter for gray-scale maps due to the fact
that the observers had to more carefully follow a gray-scale metro line which
is obviously easier if color is supported.
214 Eye Tracking
(a) (b)
Figure 5.13 Graph layouts with different kinds of link crossings, crossing angles, and the
effects of geodesic-path tendency can have varying impacts on eye movements [245].
Node-link diagrams for graphs and networks (see Figure 5.13 for
examples illustrating the ideas in the graph layout eye tracking study) can
be visualized in several layouts, all following certain aesthetic graph drawing
criteria [29, 405] with benefits for a certain well-defined graph task [323].
The eye movement data [245] can give additional insights into the task
solution strategies applied by typical graph visualization users, for example,
which impact link crossings, crossing angles, or the geodesic-path tendencies
might have on task performance and why they could be problematic. With
an iViewX head-mounted eye tracking device created by SensoMotoric
Instruments (SMI) and by recruiting 16 participants such effects of the
graph layout were investigated. As a major result it was found out that
small crossing angles can lead to increased response times and more eye
movements around the crossing areas for route finding tasks while node
location tasks are not that affected by these issues. Moreover, the criterion
of geodesic-path tendency shows that routes in a graph might be harder to
follow by the eye.
There are many more eye tracking studies taking into account
visualization techniques, for example focusing on visual variables in 3D
visualization [335] by recruiting 36 participants and by using a Tobii T120
eye tracking device. An SMI RED 250 eye tracker was used to investigate
2D and 3D visualizations in cartography [402]. Visual exploration patterns
in general information visualizations have been studied by using a Tobii
X120 eye tracker by recruiting 23 participants [5]. Scatter plots [178] and
scatter plot matrices [451] have been a focus of eye tracking. A remote Eye
Tribe eye tracker was used in the scatter plot matrix approach in which 12
participants were recruited. Ontology visualization has also been explored
5.4 Examples of Eye Tracking Studies 215
by applying eye tracking [193]. Indented list and graph approaches were
evaluated by recruiting 36 participants recording their eye movements with
a Tobii 2150 eye tracker. Even Euler diagrams have built the basis for an
eye tracking study [438] by asking 12 participants and by using a Tobii X2-
60 eye tracker. Further topics under eye tracking investigation include 2D
flow visualization [231], flow maps [157], program visualization [33], clutter
effects [358], or radial diagrams [204].
scanning are two popular viewing strategies with the goal of finding relevant
information, often resulting in an “F”-shape pattern if the web page contains
mostly textual information, as found out by other eye tracking studies [464].
However, the visual hierarchy of a web page plays a crucial role if this “F”-
shape pattern occurs or if it looks different, which also has an impact on
the way we interact with the content. Forty-eight participants were recruited
to explore the browsing strategy in web pages while the Tobii X120 eye
tracker was placed in front of a monitor. This interaction is not an active
but rather a passive interaction technique [476] in the way that the user
communicates with a stimulus to get back important information, but the
stimulus is not changed. However, it can give insights into where the user
might find information and hence needs to use scroll bars, and so it is some
kind of passive “explore” interaction [544]. The major result of this eye
tracking study is that the top-down model based on the visual hierarchy
is typically preferred, but the “F”-shape pattern of viewing also plays a
significant role.
Another interaction task, namely navigation in web pages, showed
interesting results based on the viewing behavior [543]. Eighteen participants
had to work with typical web pages and had to respond to typical navigation
tasks while they had to browse through the different linked pages to locate
a task’s result. During this scanning procedure their eye movements were
recorded with a Tobii 1750 eye tracking device. The major result of the study
is that participants tend to use reference and identification points to better
track their locations during the navigation task which is some kind of mental
map preservation to not be completely lost in such a complex task, hence
people tend to reduce their cognitive load somehow to orientate themselves
in the web page contents, in particular when the content is changing by
self-initiated mouse clicks on web page links.
Interacting with large displays plays a larger and larger role these days,
in particular in visual and immersive analytics applications. Two gaze-based
interactions for an individual user have been developed and evaluated by an
eye tracking study [283]. Walk-then-interact as well as walk-and-interact are
two different setups in the system that are investigated in the eye tracking
experiment. Twenty-six participants were recruited while the Tobii REX eye
tracker was used to measure the gazes and let the people reliably interact with
the system. The major result was that the system was in general accepted by
the users and that the interaction kick-off time was decreased to only 3.5 s,
making it pop out from the repertoire of existing solutions.
5.4 Examples of Eye Tracking Studies 217
hence they might expect to find the relevant textual information exactly at a
certain position. Eleven participants were recruited with nearly no experience
with eye tracking. The eye tracking device in use in this study was a Tobii
1720. The major results showed that left-aligned labels should be avoided.
Moreover, columns are not a good choice and should be avoided, or if this is
not possible, right-aligned labels should be used.
An eye tracking study in the medical domain investigated how medical
information is actually read, and probably understood, by patients [208].
To find insights in these aspects, 50 participants were recruited while the
EyeLink 1000 eye tracker was used to record eye movements during reading.
Major results of the study showed that there are differences in the reading and
understanding of the medical textual information compared to a simplified
variant of these texts. Hence, the eye tracking study shows that we have
to be careful with the information we provide for people who are no
domain experts, for example, in a visual analytics system in which textual
descriptions might be used in various forms.
Highlighting of text is a powerful tool to guide viewers’ visual attention
to specific relevant text fragments [121]. Such a feature is, in particular,
useful if a lot of text has to be shown that might be separated in relevant
and not that relevant information. However, the complete text has to be
shown for those readers who are not that familiar with a certain aspect while
the experienced user gets the most important parts in a highlighted fashion.
Such a highlighting effect of textual content is only possible these days due
to lots of possibilities for digital reading. An eye tracking study with six
participants showed that highlighted text areas attract the attention of people.
The highlights seem to pretend some kind of special importance. The used
eye tracking device in the study was an SMI EyeLink I.
Musical scores are some kind of special textual information that has to
be read by the musician to play successfully. The general problem with such
music sheets is the page turning which might be controlled by gaze instead of
using the musician’s hand [51]. Ten participants were recruited, all of them
with experience in music playing, in total eight pianists, one violinist, and one
euphonium player. An SMI RED500 eye tracker was used to measure, record,
and analyze the eye movements to make such a page turning functionality a
successful tool. The major result of the evaluation of the system showed that
the gaze-assisted page turning tool reduced the page navigation time by 47%,
a value that the researchers obtained by comparing to existing music score
reading systems.
220 Eye Tracking
Figure 5.14 Finding a bug in a source code typically requires to scan the whole piece of
code before one concentrates on specific parts of it [454].
5.4 Examples of Eye Tracking Studies 221
might argue that in a simple visualization technique the user tasks might be
much easier and faster to solve, for example, in a comparative user study, but
for visual analytics it is a wise decision to evaluate the system as a whole,
meaning the tasks to be solved are more explorative tasks with lots of options
to solve them and lots of system components to use.
Eye tracking is also useful for larger systems [307] compared to simple
visualization techniques. However, we must say that the recorded eye
movement data, be it for gaze-assisted interaction, to use it for detecting
design flaws, or as a recommender system supporting users at task solving,
is much more complex and is recorded over much longer time periods than
in traditional eye tracking studies investigating visualization techniques, text
reading tasks, or the effectiveness of layouts of user interfaces. A visual
analytics system might even be installed on different kinds of displays and
in varying environments; there can be a multitude of interaction techniques
incorporated, even working in a multi-modal fashion combined with gaze
interaction. Also augmented, virtual, or mixed reality techniques can be
part of a visual analytics system, typically integrated in immersive analytics
tools [347]. Moreover, additional data sources can come into play like
physiological measures, body movements, verbal feedback like think-aloud
or talk-aloud, and many more. All of these should be considered in an analysis
of the eye tracking study data to make the best of it in order to detect the
design issues and, consequently, to identify a way to improve the visual
analytics system.
Eye tracking applied to visual analytics systems is a relatively novel
concept and hence, there is not that much research focusing on the entire
visual analytics system, rather on specific components that are under
investigation from the perspective of visual attention. An eye tracking study
was conducted [379] taking into account tasks to explore networks. These
tasks were not given beforehand but the system was able to detect them
based on eye movement behavior and suggested visual adaptations. Twelve
participants were recruited while an SMI RED120 eye tracking device was
used. As a major outcome of this line of research it was found that there
seems to be some accuracy improvements for the network task of checking if
two nodes are connected.
Using scatter plot matrices for depicting multivariate data with the goal
of identifying correlations can be a challenging task, in particular if the user
of such a system is not able to focus on the most important views and data
aspects to solve given tasks. Such a problem was investigated by making use
of eye tracking [451] to support the visual exploration of scatter plot matrices
5.4 Examples of Eye Tracking Studies 225
Figure 5.15 A recommender system for scatter plot matrices equipped with eye tracking
technologies to support the data analysts [451]. Image provided by Lin Shao.
(see Figure 5.15) based on user input coming from eye movement behavior,
similar to a recommender system. 12 participants were recruited while the
Eye Tribe SDK was used to measure and transform the recorded eye tracking
data. The benefits of this experiment showed that such a system can get higher
pattern recall in comparison to a different interaction modality like mouse
input for example.
Also in the field of time-series visualization it is of particular interest to
identify temporal patterns, for example, to compare them with other patterns
on different levels of temporal granularity. A visual analysis can be a tedious
task if too many of those time-series patterns are displayed, hence some kind
of recommender system might support the viewer at those tasks [467]. Thirty
participants were involved in an eye tracking study in which an Eye Tribe
eye tracking device was applied to incorporate the eye movement data into
the data analysis and recommendation process. The evaluation of the system
showed that it is possible for the observers to quickly identify time-series
patterns that are of particular interest.
Visual analytics systems might even involve user-adaptive information
visualizations [481] that allow the configuration of a system in a user-
specified view with an adapted layout of the user interface, important views
and visualizations, active interaction techniques, and so on. Eye tracking
can help to give additional user-specific input for finding a suitable solution
for each individual user. An eye tracking study with 35 participants was
226 Eye Tracking
Table 5.2 Examples of eye tracking studies focusing on aspects in visualization, interaction,
text reading, user interface design, as well as visual analytics
Scenario Participants Eye tracker Ref.
Hierarchy visualization 40 Tobii T60 XL [78]
Trajectory visualization 25 Tobii T60 XL [369]
Public transport map 40 Tobii T60 XL [372]
Geographic map 30 Tobii T60 XL [371]
Graph visualization 16 SMI iViewX [245]
Webpage browsing 48 Tobii X120 [156]
Webpage navigation 18 Tobii 1750 [543]
Large display interaction 26 Tobii REX [283]
Plane landing 12 Smart Eye [339]
Menu selection 11 ISCAN RK726/RK520 [105]
Label positions 11 Tobii 1720 [143]
Health document reading 50 EyeLink 1000 [208]
Text highlighting 6 SMI EyeLink I [121]
Music score page turning 10 SMI RED500 [51]
Source code reading 15 Tobii 1750 [454]
UI interaction 20 SMI REDn/Tobii EyeX [353]
Multi-modal interfaces 11 Tobii x5 [36]
Item list interface 64 Tobii X2-60 [197]
Interface layout 21 Tobii 1750 [210]
Gesture-based interface 5 Tobii x2-60 [488]
Network exploration system 12 SMI RED120 [379]
SPLOM recommender system 12 Eye Tribe [451]
Time-series patterns 30 Eye Tribe [467]
User-adaptive system 35 Tobii T120 [481]
Problem solving 28 Tobii T120 [244]
conducted taking into account such challenging problems. A Tobii T120 eye
tracker was applied to predict the best possible scenario for a user, typically
focusing on the tasks to be solved by interpreting visualizations. The major
finding of this research was that there are promising initial results indicating
that the predictions made have some positive value for the system; however,
there are still a lot of open future challenges to make it a real-time adaptable
system.
Problem solving belongs to visual analytics which can occur in at
least two ways, i.e. by using interactive visualizations or by applying
algorithms for analytical problems, but in the best case, both concepts work
in combination which is actually the power of visual analytics. However,
5.4 Examples of Eye Tracking Studies 227
229
230 Eye Tracking Data Analytics
all of those to simple statistical graphics can lead to wrong conclusions [15] or
missing correlations between data dimensions that might have been visible in
a case in which more complex visualization techniques are used. Negatively,
such visualization techniques demand for learning and understanding the
new concepts [71] which is actually difficult for non-experts in visualization.
Consequently, to visually explore eye tracking data reliably we should have
some profound knowledge in eye tracking and visualization at the same time
which reduces the number of people actually using visual depictions of eye
tracking data.
Positively, standard and well-known visualization techniques like visual
attention maps [50] or gaze plots [203] which show space, time, and
participant dimensions are already well established in the eye tracking
community in a way that they are well understood and they can be found
in many results sections in scientific research papers related to eye tracking
data. Moreover, they have been so popular that they are typically built-in
features in today’s commercial eye tracking analysis software. However, from
a perceptual and visual perspective such visualizations are not as powerful as
expected. Visual attention maps aggregate over time and participants, while
gaze plots lead to visual clutter if too many scanpaths have to be shown on
a visual stimulus. These drawbacks are one major reason why many more
advanced eye tracking data visualizations have been developed [47], all trying
to show as many insights about the recorded eye tracking data as possible.
Visual analytics (see one example of a visual result in form of gaze stripes in
Figure 6.1) adds one powerful strategy to this existing repertoire of analysis
techniques [14], but again, it requires knowledge from several domains, other
than from eye tracking, to find knowledge and meaning in the eye tracking
data which is definitely one of the biggest challenges in this whole domain
combining eye tracking and visual analytics.
Figure 6.2 A manual fixation annotation tool has been developed to step-by-step add extra
information to the fixations, for example based on the semantics of a stimulus [370].
which data elements from the data source are removed, for example for data
privacy reasons. This step typically removes all personal information from the
data source, which means all data elements that might somehow be used to
recover the person behind a certain scanpath for example. A typical strategy
is to replace all person names with identifiers that are just random natural
numbers. It should not be possible to link these identification numbers with
the filled out study participant forms to definitely avoid recovering personal
information like the person’s name. A problem is definitely the fact that
certain person’s names might be recovered in the case that a certain attribute
value only exists once and hence can be mapped uniquely to this person. An
example would be if we just had one person who is wearing contact lenses,
then this person can be identified easily later on in the recorded scanpath
data by just looking up this special attribute. The safest anonymization
effect in an eye tracking study is to completely avoid the recording of a
person’s name, but just in case the persons are paid for participation in an
eye tracking experiment we normally need the names. However, this extra
payment information should be separated from the recorded data and maybe
should be destroyed as soon as possible to avoid anonymization problems
later on.
tracking data sources separately or to transform each data source into the
same format based on the same template.
If several data sources are stored for a later data analysis, this problem
might get even worse. For example, the scanpaths from an eye tracking
study as well as additional verbal feedback, interview data, personal
study participant data, physiological measures, and so on might be worth
investigating. All of those data sources could be stored in different formats,
for example, stemming from various eye tracking experiments conducted by
different groups of researchers. Finding a data analysis tool that helps to
identify data patterns and insights in those data sources is a difficult task
since, in the worst case, each data scenario has to be adapted to the tool’s
data reading and parsing requirements. If this adaptation happens on the tool
side it can be challenging for the tool developer to create a parsing function
for each of the data formats. Maybe only the most important ones might be
supported by a tool, the rest has to be brought in the correct template by the
research group who generated a dataset. This is typically the best option since
they are best in interpreting their own data and deciding which values in the
dataset belong to which attributes for example.
time. Hence, the linking does not only happen between several separate data
sources, but also over time, in case we have to deal with dynamic visual
stimuli, which is a typical scenario in visual analytics.
The general problem might occur if the data sources have to be linked
during the runtime of a data analysis or visual analytics tool. For example,
a user might decide to visually explore eye movement data by also taking
into account an additional data source describing interview data that has been
acquired by asking the study participants after each of the experiments. Then
this extra data has to be linked with the primary eye movement data first and
then additional views or visual output have to be incorporated into the original
visualization for example. Although the visualization of the extra data might
not be a problem at all, it might take some time to link the two data sources
before the visualization can be modified and updated. Even if such a linking
process just takes a few seconds it might even lead to a bad user experience.
Consequently, the question comes up which data sources should be linked
beforehand, i.e. before the analysis tool is running to avoid such waiting times
during the data analysis process. Moreover, another question is how the data
sources can be linked, meaning if there is some common key that can be
used to do that successfully, for example, as in our scenario above, the study
participant might be the key to link the data sources. In some situations it is
not clear how the linking can be done reliably without asking the user of a
data analysis tool which makes the data linking a quite challenging topic.
this early stage while it is stored or added to the general eye tracking database
for later usage.
Important stages in this whole process include the checking of the data for
inconsistencies or redundancies. Consequently, the data sources get validated,
verified, and typically cleaned and freed from errors or missing data entries
whenever this is possible. This builds some kind of data enhancement or
data enrichment process in which certain data elements can be removed,
added, or even annotated by special events or uncertainty values to indicate
that they might otherwise lead to misinterpretations. Moreover, in this step,
additional analysis-relevant data metrics can be derived and values for them
can be generated. Computing those values beforehand can save a lot of
computation time during the running data analysis system. Hence, this whole
stage focuses on storing the data in an efficient way, on adapting it in a
way that it can be accessed as quickly as possible later on, to save valuable
computation time during a running data analysis or visual analytics system,
and finally, on allowing data transformations that might modify a raw dataset
into a computer-readable one, for example, based on bringing it into a given
format that is understandable by a data analytics tool. This kind of data
transformation is just responsible for adapting the format of the data, it
does not draw any conclusions from the data nor does it derive any data
patterns, like an explicit algorithmic data analysis process would do which
will be discussed in one of the following sections. Generally spoken, when
a visual analytics system for eye tracking data is started, the underlying data
to be analyzed should be in the most appropriate data format as possible,
for example, to avoid long runtimes while using the visual analytics tool and
while interacting with it. Nobody wants to wait too long for the results of a
data analysis, but in some situations it is unavoidable.
matter how efficient the data is stored in these scenarios. For eye tracking
data it also depends on which data dimension the data analysts are primarily
interested in, like the spatial information from the visual stimulus, the
temporal information, i.e. how the situation changed over time, for example,
the stimulus and/or the eye movement behavior, or the eye tracking study
participants, individually or clustered into participant groups. The primary
data dimension typically decides which algorithms and visualizations are
used later on and hence the data should be stored in a way that this primary
aspect can be accessed as fast as possible while the secondary or tertiary data
dimension plays a minor role.
For example, if a visual analytics system focuses on supporting an
analysis of participant clusters, the individual participants are of primary
interest and not the visual stimuli. However, the stimuli might be used as a
later details-on-demand request to see where certain participant clusters paid
visual attention, but actually for identifying the participant cluster this data
is not as important as the participants themselves, maybe with their personal
information. Things are not that easy in most situations. To decide on the
order of relevance of the data dimensions and how to structure and organize
them is a difficult problem. On the one hand it definitely depends on the
data analysis tool or visual analytics system that works with this kind of
data, but on the other hand it also strongly depends on the user tasks. From
the scenario above, it might be efficiently stored if the major focus is on
participant clusters based on personal information, but if the users decide to
switch to a more scanpath- and stimulus-related grouping of the participants
instead of the personal information, the performance of the system might
suffer from it.
Today, with only small eye movement datasets, we might argue that this
is actually not a big challenge, but in future scenarios, for example, with
millions of scanpaths and dynamic stimuli of hours of lengths, we might get
into serious performance issues if this problem is not treated well enough.
One scenario could be the tracking of the eyes of millions of car drivers
over longer driving distances. On the one hand, we would like to analyze
the scanpaths after the driving tasks have been finished which actually gives
us enough time to process the data, i.e. making the storage problem actually
not that relevant. However, if we are facing a scenario in which we already
have millions of scanpaths from car drivers and now a new car driver is eye
tracked, we might want to get real-time feedback based on the eye movements
of the one car driver while at the same time taking into account the existing
scanpaths, typically stored in a database, prepared for such purposes. There
240 Eye Tracking Data Analytics
are many of such future scenarios in which we might face this challenging
problem, the larger the eye tracking data sources get, the more insights we
can extract from them, but at the cost of thinking about efficiently storing and
managing the data.
process for low-level eye tracking data [444]. Positively, additional data
sources might be a good option to more efficiently and reliably clean the data
since they allow us to take into account further data perspectives compared to
just one individual data source alone.
the temporal granularity of the data. This could be very relevant information
if the recorded data stems from different eye tracking devices with varying
properties. Consequently, the type of the eye tracker should also be stored as
well as additional meta information about the experiment.
deciding during the runtime of the system which parts of the data to be
transformed might lead to a performance degradation and, consequently, it
is a wise decision to preprocess the data and transform everything that can be
transformed while taking into account the fact that after the transformation
the granularity of the data is changed. This means if we aggregate or even
normalize the given recorded raw data beforehand, for example, changing
from a lower to a higher scale, we no longer see the lower scale in the data
analysis or visual analytics tool. One solution could be to still keep the raw
data, just in case the tool user makes a request to it which might lead to
runtime performance issues, but in the case that the data is used on the higher
scale, the tool might perform quite well.
is, on the one hand, an offline approach that starts computing after the data
has been recorded completely, i.e. as some kind of post-processing. The other
alternative might be denoted by the term online approach which indicates that
the data has to be algorithmically analyzed during the recording, i.e. in real-
time. The second option is typically more complicated to implement since
it requires the algorithm to react quickly, i.e. in real-time, on a given input.
This could be interesting for gaze-assisted interaction for which the data of
many eye tracked people might be taken into consideration to generate a
quick recommendation while interacting. Moreover, in any scenario in which
the user dynamically interacts or inspects a visual stimulus, an algorithm
might generate real-time solutions and hints during this dynamic process.
The offline approach, on the other hand, is typically useful when we have
enough time to analyze the data, for example, in an eye tracking study for
which we have recorded all the data beforehand. As a post-process, i.e. a
data evaluation and analysis, we might apply a visual analytics system with
various algorithms [14], to find design flaws in the given stimuli during the
eye tracking study.
Figure 6.3 The Antwerp public transport map was visually explored in an eye tracking study.
The visual attention hot spots were used to split the static stimulus into sub-images which are
then grouped by a force-directed layout taking into account the transition frequencies between
the individual sub-images [98]. Different parameters can be modified such as cropping sizes,
cluster radius, or number of sub-images displayed, for example.
the different viewing behaviors that cause the changes in the scanpath data.
It may be noted that the longer the scanpaths are under investigation, then
the probability that they share common data patterns is normally lower.
Moreover, there are always small variations in the fixation pattern for each
scanpath, even if they are very similar. This problem might be mitigated by
either using spatio-temporal thresholds for each fixation that still consider
fixation points as being similar if they are located within the same radius, for
example. Another idea to manually or automatically annotate or manipulate
each scanpath is based on the semantics of a stimulus, hence identifying
the visual object a fixation belongs to in a stimulus. However, this can be
a time-consuming process, in particular, if it is done manually [370].
interest they create during the visual scanning strategy. Hence, such a classing
and classification might be a useful concept to guide a clustering, for example
a hierarchical one, taking into account the participants and their scanning
behavior reduced to certain classes based on well-defined properties. On the
one hand, we lose information by the classing, while on the other hand, the
follow-up clustering algorithms might produce faster results since the input
data is no longer that accurate, but it is condensed to the most relevant aspects
that are still detailed enough to generate a suitable and expressive clustering,
for example. The classing actually works for any data dimension, we only
have to decide how the classes are created and how many we plan to get
in the end to base further data organization, structuring, or clustering on.
Moreover, if another scanpath is recorded it is faster and easier to classify to
which category it belongs than taking into account the whole scanpath with
all its detailed fixations and fixation duration. In particular, in the field of
fatigue detection of car drivers based on eye tracking [285] we can find many
approaches making use of classification concepts, an approach that typically
requires real-time computations to provide fast results.
format that allows better comparisons [230]. Choosing the logarithm idea
actually transforms the quantities to exponents, allowing the small values to
be compared visually. However, the users of such a logarithmic scale should
be informed about this data transformation to avoid misinterpretations.
Aggregation, as mentioned earlier, is also a useful concept to reduce the
dataset size and to provide a better overview of the data, in particular, if
the data reaches sizes that no longer allow it to be visually depicted without
needing the user to scroll a lot. In some scenarios we can find a combination
of these concepts, i.e. the eye movement data is aggregated, normalized, and
finally, transformed to a different scale. All of this follows the goal to get a
scalable overview of the eye movement data for as many data dimensions as
possible to allow fast comparisons. This first overview strategy is a fruitful
concept to support a data analyst with a starting point for further exploration
processes, algorithmically as well as visually. The aggregation strategy can
be manifold and could be based on several approaches, including simple and
more complex statistical ones. For example, if we plan to inspect the fixation
durations in a scanpath and plan to temporally aggregate those, we may ask
the question what the result of such an aggregation strategy might be. We
could generate the sum of the fixation durations but without normalizing the
sum by taking into account the number of fixations in a time interval, this
approach might be misleading, hence the average or mean value might be of
special interest here. Also the minimum valley or maximum peak might be
interesting aggregation measures for each time period. More statistical values
such as the median or the standard deviation could give even more detail.
Also a combination of several of those aggregation measures could be useful,
in particular if those are visually depicted later on, for example in box plots.
one attribute also increases the value of another or even more other attributes.
In contrast, the increase could also lead to a decrease for other attributes.
The first observation is called a positive correlation while the second kind
is denoted negative correlation. Not only the static correlation behavior of
attribute values might be of interest but also the dynamic ones, i.e. it could be
of particular value to identify the correlation behavior over time between two
or even more attributes. For example, the average fixation duration in a certain
time period might correlate in a specific way with the average saccade length
in the same time period. If this time period is moved just like a sliding time
window over the entire scanpath we could analyze if the dynamic correlation
pattern changes or stays the same. The scanpath with the attributes fixation
duration and saccade length is just a simple example for such a correlation
analysis but there is no limitation to extend it to any kind of quantitative
attribute, static or dynamic. The quantitative eye movement metrics could
even be set in correlation to other metrics not directly related to the eye [540].
If the dynamics in the eye tracking data plays a crucial role for further
investigations we might consider trend analyses [175]. These algorithmic
approaches take into account a time-varying dataset and compute a trend
in it based on certain attributes and data properties. For example, for eye
movement data it might be of interest to analyze how the fixation duration
changes over time since there is some kind of evidence that the fixation
duration can give hints about certain task solution strategies and how much
effort a viewer is putting into a certain task. A similar aspect holds for the
saccade lengths, i.e. analyzing the saccade lengths can also give insights into
the modification of a certain viewing behavior or scanning strategy. Trend
analysis can help to uncover increasing or decreasing effects in a dynamic
dataset, but also constant behavior, oscillating or alternating effects, as well
as outliers and anomalies. Moreover, considering correlations we might even
combine trend analysis with correlation analysis to identify countertrends in
a dynamic dataset, for example, an increasing trend pattern might hold for
one attribute, but compared to another one which shows a decreasing trend
pattern, the combination of both patterns would uncover a countertrend, i.e.
one dynamic attribute shows an opposite effect compared to another one
or even many more. The trend analysis might even be carried out for an
individual attribute, like the saccade length over time, and then if it is applied
to all scanpaths of all eye tracking study participants, those trend detection
results might even be usable for grouping participants. However, the grouping
strongly depends on the attribute under investigation and also how long the
corresponding scanpaths are, for example.
252 Eye Tracking Data Analytics
Figure 6.4 Alignment of a set of scanpaths from an eye tracking study. First, the scanpaths
are transformed into character sequences based on user input, before they are aligned [84].
computed rules to new situations. The learning can happen in different ways
like supervised, semi-supervised, unsupervised, or as reinforcement. Multi-
layered neural networks [384] are often used to train, as a mechanism to
perform complex tasks in larger and larger datasets to which eye tracking
data also belongs. Although artificial intelligence has generated various fast,
efficient, and quite accurate methods, the whole discipline is just about to start
to take into account eye tracking data. The major reason is that the available
eye tracking data today might still be too small to make reliable predictions
based on artificial intelligence, using machine and deep learning approaches.
However, a few problems have been tackled in the field of eye tracking
by making use of AI-related concepts. For example, convolutional neural
networks have been used for analyzing real-time eye tracking data with focus
on interactive applications [77]. In most scenarios the research focuses on
eye images, for example, to train a machine learning algorithm based on a
multitude of such images to classify or predict newly seen images. Such an
approach is, in particular, useful for detecting negative performance issues
of car or truck drivers, for example, if their eye movements indicate fatigue
effects that might cause accidents [115]. Such image-centric tasks are hard to
solve by standard algorithms due to the vast amount of data and features to be
explored, in particular, if real-time analyses are required. Machine learning,
on the other hand, can be used to make fast predictions and classifications,
however a certain large amount of training data is required to generate
accurate and meaningful results. For example, a model for predicting where
people look in images [302], predicting gaze fixations [133], or saliency in
context to predict visual attention [249] are typical research areas.
least three major data dimensions which come in the form of space, time,
and participants in an eye tracking study. Depending on the tasks [304] the
users of a visualization technique or visual analytics system plan to solve,
the visual encoding can vary a lot as well as the interaction techniques that
are integrated into the provided visual depictions of the data. Moreover, the
way in which the data actually exists or is being transformed plays a crucial
role for the visual metaphor and the visual variables in use. For example, it
makes a difference for the visualization technique whether we are interested
in the raw fixations to a stimulus or to spatially aggregated areas of interest,
while the dynamics of the data plays a role for the visual depiction, static
data is definitely easier to visualize than its time-varying variant, i.e. several
instances of the static data.
Visualization techniques for eye tracking data exist in various forms [47],
either focusing on individual aspects in the data, or incorporating more and
more data dimensions and derived values, typically attached to one or more
of the provided visualization techniques focusing on the primary aspect in
the data based on the primary task or tasks a user wishes to solve with a
visualization, or at least get a hint about a certain visual pattern that initiates
further exploration and analysis processes. Visual analytics goes one step
further than traditional visualization techniques since it is an interdisciplinary
field that combines concepts from algorithmics, statistics, human–computer
interaction, visualization, perception, cognitive processing, and many more.
Hence, with visual analytics we can actually get power from both sides,
the machine and the human side, to build models and hypotheses for our
eye tracking data guided by interactive visual depictions of the interesting
pieces of the data, finally leading to insights and knowledge from those
large and heterogeneous eye tracking data sources, in particular in future
scenarios when eye tracking data grows and grows [44] with many more
extra data sources about human behavior and additional personal feedback.
Visualization and visual analytics are not built to solve the problems in eye
tracking data, but in cases where the data is visually encoded in a perceptually
and visually effective way, we can recognize visual patterns that can be
remapped to data patterns in the best case, leading to the formulation of new
hypotheses and also to the confirmation, rejection, or refinement of already
existing hypotheses. Visualization plays the role of guide through our large
eye tracking data since it allows us to navigate, scroll, filter, and finally,
explore the data.
256 Eye Tracking Data Analytics
all the values as a bar separated by the median line that reflects the value
exactly in the middle of the distribution. Box plots can be useful to show
fixation deviations with the goal of analyzing the deterioration of the eye
tracking device calibration [240] or as a summarized comparison of study
participants and their normalized scanpath saliency scores [158]. Analyzing
pairwise correlations can be done with scatterplots while each attribute is
mapped to one of the axes. Also, in the field of eye tracking, scatter plots have
been used, for example, to plot correlations between response latency and
angular disparity [268]. Moreover, scatter plots can be used for comparing
different species’ eye movements like humans and monkeys, for example, by
plotting amplitudes and velocities of the recorded saccades [35]. To explore
more than two attributes we might use star plots, for example, to visualize
scanpath properties [203] or fixations [365]. Apart from star plots we can use
parallel coordinate plots to show correlations between several attributes like
derived eye tracking metrics [299].
effects and are difficult to interpret because of missing reference points to the
axes. As a negative consequence of splitting the x- and y-coordinates of a
scanpath to show them separately in timeline plots we cannot easily identify
the temporal changes in the spatial dimension, i.e. in the visual stimulus.
To see this effect we can use the popular gaze plots [448], however, if the
scanpaths are quite long or many participants’ scanpaths have to be shown at
the same time, we reach a problem denoted by visual clutter [426].
Instead of showing the fixation points over time we might be interested
in inspecting the fixation data with an additional view on the shown stimulus,
i.e. the spatial dimension. This could be done using a transparent overlay on
the stimulus, however, the dynamics of the fixation data can also be shown,
for example, by an animated diagram known as a bee swarm visualization [1],
also for a dynamic stimulus [345]. However, animation is typically considered
to be problematic for comparisons over time [505] due to the fact that a
viewer has to remember lots of visual patterns in the short-term memory to
reliably do the comparisons. A static side-by-side visualization might be the
better option for such data although the display space is a limitation of the
static representation. If the fixation data is temporally aggregated as well as
over groups of participants we denote such a visualization as visual attention
map, fixation map, or heat map [49, 50, 473]. This aggregated representation
of the fixation data does not show the time-varying behavior, but serves
as a great overview of visual attention hot spots—regions in the visual
stimulus that attracted much attention. Such hot spot regions are typically
used for defining areas of interest (see Section 6.4.3). The visual depiction of
these hot spots can be based on several criteria like fixation count, fixation
duration, relative fixation duration, or participant percentage, to mention a
few. Moreover, the visual appearance of the attention maps can be based on
several visual variables (see Figure 6.7 for visual attention maps enhanced
by contour lines); typically color coding [50, 167] is used to visually depict
6.4 Visualization Techniques and Visual Analytics 259
(a) (b)
Figure 6.7 Two different visual attention maps from a public transport map eye tracking
study. In this case the hot spots of visual attention are indicated by contour lines [100]. Route
finding tasks in the maps of: (a) Tokyo, Japan; (b) Hamburg, Germany.
the visual attention at a certain point, but also luminance [515], contour
lines [100, 158, 206], or even 3D effects [321, 530]. However, although
attention maps seem to be powerful concepts they also have to be taken with
care [235] to avoid misinterpretations of the data. For a dynamically changing
stimulus it is challenging to generate a visual attention map. However,
synchronous attention of several participants can be computed and then this
can be visualized over time [357]. Also motion-compensated attention maps
based on optical flow concepts for moving objects have been explored [311].
On the other hand, for 3D visual stimuli, the 3D visual attention is typically
directly incorporated in the stimulus [394], maybe also by 2D projections
or by coloring the entire 3D visual object by using the visual attention map
color [483].
If we are interested in seeing the connected fixations, i.e. the whole
scanpath, over space and time we can use scanpath visualizations [375],
meaning a sequence of fixations and saccades [235]. A standard visualization
for a scanpath is a composition of circles of different size with the circle
center where the fixation was done in the visual stimulus, encoding the
fixation duration in the circle size while the saccades are represented
as straight lines connecting subsequent fixations [448]. The scanpath
visualization is typically overplotting the visual stimulus for contextual
reasons (see Figure 6.8 for examples of scanpath visualizations). Also the
velocity of fixation data can be encoded [318] while in the early days the
fixations were not visually indicated, just the saccades [539]. Scanpaths can
even be used to derive other visual effects, for example the convex hull
260 Eye Tracking Data Analytics
(a) (b)
Figure 6.8 Scanpath visualizations for (a) one participant and (b) 40 participants [100]. The
scanpath visualization in (b) can hardly by used for data exploration.
Figure 6.9 A space-time cube showing clustered gaze data for a given stimulus [311]. Image
provided by Kuno Kurzhals.
(a) (b)
Figure 6.10 AOI visits over time, either for one participant and three AOIs [414] (a) or
three participants and four AOIs in parallel [417] (b). Extending these visualizations to many
participants, many AOIs, and long scanpaths can lead to visual clutter effects.
(a) (b)
Figure 6.11 Annotating a visual stimulus, overplotted with a contour visual attention map,
with color coded AOIs (a); the AOI visits over time can be seen in a corresponding scarf plot
(b) that uses the same color coding as in the annotation view.
Figure 6.12 The dynamic AOI transitions can be shown in an AOI river visualization [80]
with an enhancement by Voronoi cells.
combination of graph and hierarchical aspects among the AOIs can be used to
compute a hierarchical graph layout for the AOIs together with their transition
frequencies [81].
building of models and visualizing them in a way that users can step-by-step
explore their eye tracking data. To reach this goal, visual analytics applies
techniques from many application fields [277] due to the fact that it is an
interdisciplinary approach. However, visual analytics cannot solve all eye
tracking exploration problems [412], but at least it provides ideas that show
how to come closer to solutions, in cases where the data problem seems to
be algorithmically intractable or the required visualization produces a non-
scalable representation of the data. For example, real-time eye tracking data
analysis [482] is a challenging problem that needs powerful concepts from
various disciplines to keep up with the flood of data that is recorded, even
from several eye tracking devices at the same time, maybe generated in
scenarios involving VR/AR/MR or in particular, immersive analytics.
Gaze stripes [309] or ISeeCube [308] focus on visual analytics of video
data by providing visual and algorithmic concepts in combination. Both
visual analytics concepts can also be modified to make them applicable
to static visual stimuli. ISeeColor [390] combines interactive visualizations
and automatic recognition of independent objects by applying deep learning
approaches focusing on semantic segmentation. Also the identification of
reading patterns is in focus of eye tracking visual analytics [538]. Making
a visual analytics tool for eye tracking accessible via the internet (see
Figure 6.13) is a good idea to reach many users and to get feedback [22].
Figure 6.13 A graphical user interface showing several linked views for visually exploring
eye movement data: a clustered fixation-based visual attention map, a timeline view on the
visually attended AOIs, a scanpath visualization, a visual attention map with color coded hot
spots, and a scarf plot for an overview about the inspected AOIs [22].
6.4 Visualization Techniques and Visual Analytics 265
Moreover, researchers can upload and share their eye tracking data with
others, as well as the found insights. ETGraph is a system for eye tracking
visual analytics based on graphs [211]. Also in the medical domain there are
tools making use of eye tracking data and trying to analyze them with visual
analytics concepts [472].
7
Open Challenges, Problems, and Difficulties
267
268 Open Challenges, Problems, and Difficulties
problems that lead to negative aspects concerning the recorded data. The eye
movement data can be inaccurate in many ways, making it hard to use or
to make predictions based on the users’ visual scanning behavior. Hence,
it is a good advice to check the reliability of the recorded data before it is
analyzed and visualized to avoid wrong conclusions drawn from it. No matter
how accurate the data is, the eye-mind hypothesis leads to the fact that the
data is regarded as useful or not for describing what people are cognitively
processing while they are visually inspecting a static or dynamic stimulus
which also makes the interpretation of eye tracking data a challenging field
of research. The eye–mind hypothesis can have a negative impact for the
whole eye tracking data analysis and visualization community, depending on
whether we believe in it or not.
Another eye tracking challenge is caused by the costs that come with each
eye tracking device [44]. Although those costs have been decreasing over
the years due to the progress in hardware technology like faster processor
speed or improved digital video processing and the fact that the market for
selling those devices is growing a lot, it can still be quite expensive, but this
typically depends on the application domain and parameters like accuracy
and tracking rate, i.e. all the required technologies involved in building a
suitable eye tracking device. Moreover, nowadays there is some kind of
competition between eye tracking companies (see Section 5.2.4) trying to
design the best solution for any kind of application scenario, a fact that can
also lead to cheaper devices, but mostly for general applications using the
standard devices. One such emerging field of eye tracking research focuses
on web-based or online eye tracking studies that are particularly useful in
pandemic times such as those we are facing at the moment. However, online
studies are typically uncontrolled. For eye tracking studies, this aspect also
means that each participant must be equipped with an eye tracker or the eye
movement data has to be recorded by other novel devices, for example, by
using a webcam which, on the negative and challenging side, does not allow
producing quite accurate eye movement data over space and time. This is also
one of the challenging difficulties for eye tracking integrated in smartphones,
either for analyzing the design of an app or for using it in the mode of gaze-
assisted interaction to modify something in a user interface or to navigate in
an app.
Allowing eye tracking to be used in smartphones might make this
technology applicable in a collaborative manner, i.e. making use of the
scanpaths of several people in different places in the world, but this powerful
idea causes even further challenges related to data privacy issues and ethical
7.2 Eye Tracking Visual Analytics Challenges 269
bigger and bigger, stemming from several heterogeneous data sources creates
more and more challenges for all of the data analytics and visualization
fields. Moreover, real-time analyses require the most powerful concepts to
keep up with the pace of the growing eye tracking datasets. For example,
having recorded various scanpaths beforehand and trying to analyze a new
incoming scanpath in the light of the existing data might be a great idea,
but to achieve fast solutions, i.e. in real-time, we need advanced algorithmic
approaches that can efficiently tackle such data scenarios. Fields like artificial
intelligence, machine learning, deep learning, data mining, and the like play
more and more important roles in the data analysis, however visualization and
visual analytics are suitable concepts to involve the human users with their
perceptual and visual strengths. But negatively, the human users are typically
not able to solve real-time data analysis problems, they can more or less guide
the analysis process on a visual exploration basis. For example, human users
can decide which kinds of algorithms to apply to a certain data problem or
they can include additional information in the data analysis process, maybe
the semantics given in a visual stimulus which is something that an algorithm
can hardly involve automatically in the analysis, unless it is not trained with
various stimuli beforehand.
For a visual analytics system it might be a challenge to reliably connect
itself with the recorded eye movement data. This is in particular problematic
if several eye tracking devices are used or a new one that has not been used
before comes into play which requires first adapting to the new data format.
Moreover, it is unclear if the visual analytics system is able to keep up with
the growing dataset sizes, in particular if dynamic stimuli like videos are
included. It is also unclear if such a system can handle both types of stimuli,
static and dynamic ones, also those with actively changeable content like
interactive user interfaces. The scalability issues might arise if eye tracking
is integrated in smartphones one day, producing vast amounts of data worth
analyzing, even in real-time which might cause further problems for storing
the data, transforming it, and finally, accessing it again to make predictions
or recommendations based on the formerly recorded data. Also the display
has a crucial role for a visual analytics system, meaning small-, medium-,
or large-sized displays, all having their benefits and drawbacks. Those are
typical problems that come with growing dataset sizes including research
from the field of big data [44], involving various disciplines to analyze the
data with the goal to detect rules, correlations, patterns, and finally, insights
and knowledge, even in real-time. Another aspect from the perspective of
visual analytics systems could be the idea of letting users interact with the
7.2 Eye Tracking Visual Analytics Challenges 271
system while at the same time their eye movements are recorded. This data
can be analyzed while users further work with the system and, based on the
outcomes of such an analysis, the visual analytics system might be adapted
to some degree, making the whole concept some kind of dynamic visual
analytics system based on user behavior like visual scanning strategies.
Many application fields could benefit from eye tracking as well as visual
analytics in combination. For example, the field of medicine might apply eye
tracking to later analyze and understand how doctors behaved during surgery,
i.e. where they looked over time. Moreover, a similar scenario holds for
aircraft pilots who first control a plane, for example, in a landing maneuver.
Recording the eye movements and analyzing and visualizing the data later
on can help to identify which visual elements the pilot has missed during
the maneuver. This again might help to improve the landing strategy for later
training phases. Such insights could be helpful to better train young doctors
or pilots to get more practice for future surgeries or landing maneuvers.
Education in eye tracking as well as visual analytics [71] is a deciding factor
to train young researchers and to make them aware of the challenges in both
fields, but also the benefits and synergy effects that might come with such
a combination. On the negative side, it is quite difficult to educate young
students since the fields of eye tracking and visual analytics both involve that
many concepts that it is impossible to teach these topics in a short time period,
hence only the tip of the iceberg can be the focus of education.
References
[1] TOBII TECHNOLOGY AB. Tobii studio 1.x user manual, 2008.
[2] Moataz Abdelaal, Marcel Hlawatsch, Michael Burch, and Daniel
Weiskopf. Clustering for stacked edge splatting. In Fabian Beck,
Carsten Dachsbacher, and Filip Sadlo, editors, Proceedings of 23rd
International Symposium on Vision, Modeling, and Visualization,
VMV, pages 127–134. Eurographics Association, 2018.
[3] Athos Agapiou. Remote sensing heritage in a petabyte-scale: satellite
data and heritage earth engine© applications. International Journal of
Digital Earth, 10(1):85–102, 2017.
[4] Wolfgang Aigner, Silvia Miksch, Heidrun Schumann, and Christian
Tominski. Visualization of Time-Oriented Data. Human-Computer
Interaction Series. Springer, 2011.
[5] Jumana Almahmoud, Saleh Albeaik, Tarfah Alrashed, and Almaha
Almalki. Visual exploration patterns in information visualizations:
Insights from eye tracking. In Gabriele Meiselwitz, editor, Proceedings
of International Conference on Social Computing and Social Media.
Applications and Analytics, volume 10283 of Lecture Notes in
Computer Science, pages 357–366. Springer, 2017.
[6] Basak Alper, Benjamin Bach, Nathalie Henry Riche, Tobias Isenberg,
and Jean-Daniel Fekete. Weighted graph comparison techniques
for brain connectivity analysis. In Proceedings of ACM SIGCHI
Conference on Human Factors in Computing Systems, pages 483–492,
2013.
[7] Mohammad Alsaffar, Lyn Pemberton, Karina Rodriguez-Echavarria,
and Mithileysh Sathiyanarayanan. Visual behaviour in searching
information: A preliminary eye tracking study. In Saïd Assar, Oscar
Pastor, and Haralambos Mouratidis, editors, Proceedings of 11th
International Conference on Research Challenges in Information
Science, RCIS, pages 365–370. IEEE, 2017.
273
274 References
[96] Michael Burch, Yves Staudt, Sina Frommer, Janis Uttenweiler, Peter
Grupp, Steffen Hähnle, Josia Scheytt, and Uwe Kloos. Pasvis:
enhancing public transport maps with interactive passenger data
visualizations. In Michael Burch, Michel A. Westenberg, Quang Vinh
Nguyen, and Ying Zhao, editors, Proceedings of the 13th International
Symposium on Visual Information Communication and Interaction,
VINCI, pages 13:1–13:8. ACM, 2020.
[97] Michael Burch, Corinna Vehlow, Natalia Konevtsova, and Daniel
Weiskopf. Evaluating partially drawn links for directed graph edges.
In Marc J. van Kreveld and Bettina Speckmann, editors, Proceedings
of the 19th International Symposium on Graph Drawing, GD, volume
7034 of Lecture Notes in Computer Science, pages 226–237. Springer,
2011.
[98] Michael Burch, Alberto Veneri, and Bangjie Sun. Exploring eye
movement data with image-based clustering. Journal of Visualization,
23(4):677–694, 2020.
[99] Michael Burch, Adrian Vramulet, Alex Thieme, Alina Vorobiova,
Denis Shehu, Mara Miulescu, Mehrdad Farsadyar, and Tar van
Krieken. Vizwick: a multiperspective view of hierarchical data. In
Michael Burch, Michel A. Westenberg, Quang Vinh Nguyen, and
Ying Zhao, editors, Proceedings of the 13th International Symposium
on Visual Information Communication and Interaction, VINCI, pages
23:1–23:5. ACM, 2020.
[100] Michael Burch, Günter Wallner, Nick Broeks, Lulof Piree, Nynke
Boonstra, Paul Vlaswinkel, Silke Franken, and Vince van Wijk. The
power of linked eye movement data visualization. In Proceedings of the
Symposium on Eye Tracking Research and Applications, ETRA. ACM,
2021.
[101] Michael Burch and Daniel Weiskopf. Visualizing dynamic quantitative
data in hierarchies - timeedgetrees: Attaching dynamic weights
to tree edges. In Gabriela Csurka, Martin Kraus, and José Braz,
editors, Proceedings of the International Conference on Imaging
Theory and Applications and International Conference on Information
Visualization Theory and Applications, pages 177–186. SciTePress,
2011.
[102] Michael Burch and Daniel Weiskopf. On the Benefits and Drawbacks
of Radial Diagrams, pages 429–451. Springer New York, New York,
NY, 2014.
References 285
[103] Michael Burch, Robin Woods, Rudolf Netzel, and Daniel Weiskopf.
The challenges of designing metro maps. In Nadia Magnenat-
Thalmann, Paul Richard, Lars Linsen, Alexandru C. Telea, Sebastiano
Battiato, Francisco H. Imai, and José Braz, editors, Proceedings of the
11th Joint Conference on Computer Vision, Imaging and Computer
Graphics Theory and Applications, (VISIGRAPP, pages 197–204.
SciTePress, 2016.
[104] Wolfgang Büschel, Patrick Reipschläger, Ricardo Langner, and
Raimund Dachselt. Investigating the use of spatial interaction for
3d data visualization on mobile devices. In Sriram Subramanian,
Jürgen Steimle, Raimund Dachselt, Diego Martínez Plasencia, and
Tovi Grossman, editors, Proceedings of the Interactive Surfaces and
Spaces, ISS, pages 62–71. ACM, 2017.
[105] Michael D. Byrne, John R. Anderson, Scott Douglass, and Michael
Matessa. Eye tracking the visual search of click-down menus. In
Marian G. Williams and Mark W. Altom, editors, Proceedings of
the CHI Conference on Human Factors in Computing Systems, pages
402–409. ACM, 1999.
[106] Bram C. M. Cappers, Paulus N. Meessen, Sandro Etalle, and Jarke J.
van Wijk. Eventpad: Rapid malware analysis and reverse engineering
using visual analytics. In Diane Staheli, Celeste Lyn Paul, Jörn
Kohlhammer, Daniel M. Best, Stoney Trent, Nicolas Prigent, Robert
Gove, and Graig Sauer, editors, Proceedings of IEEE Symposium on
Visualization for Cyber Security, VizSec, pages 1–8. IEEE, 2018.
[107] Stuart Card. Information visualization, in A. Sears and J.A. Jacko
(eds.), The Human-Computer Interaction Handbook: Fundamentals,
Evolving Technologies, and Emerging Applications. Lawrence
Erlbaum Assoc Inc, 2007.
[108] Stuart K. Card, Jock D. Mackinlay, and Ben Shneiderman. Readings
in information visualization - using vision to think. Academic Press,
1999.
[109] Stuart K. Card, Thomas P. Moran, and Allen Newell. The psychology
of human-computer interaction. Erlbaum, 1983.
[110] Jorge C. S. Cardoso. Comparison of gesture, gamepad, and gaze-
based locomotion for VR worlds. In Dieter Kranzlmüller and Gudrun
Klinker, editors, Proceedings of the 22nd ACM Conference on Virtual
Reality Software and Technology, VRST, pages 319–320. ACM, 2016.
[111] Sheelagh Carpendale. Evaluating information visualizations. In
Andreas Kerren, John T. Stasko, Jean-Daniel Fekete, and Chris
286 References
[119] Shiwei Cheng and Ying Liu. Eye-tracking based adaptive user
interface: implicit human-computer interaction for preference
indication. Journal on Multimodal User Interfaces, 5(1-2):77–84,
2012.
[120] Dimitri A. Chernyak and Lawrence W. Stark. Top-down guided
eye movements: peripheral model. In Bernice E. Rogowitz and
Thrasyvoulos N. Pappas, editors, Proceedings of Human Vision and
Electronic Imaging VI, volume 4299 of SPIE Proceedings, pages
349–360. SPIE, 2001.
[121] Ed Huai-hsin Chi, Michelle Gumbrecht, and Lichan Hong. Visual
foraging of highlighted text: An eye-tracking study. In Julie A. Jacko,
editor, Proceedings of 12th International Conference on Intelligent
Multimodal Interaction Environments, volume 4552 of Lecture Notes
in Computer Science, pages 589–598. Springer, 2007.
[122] Kang-A. Choi, Chunfei Ma, and Sung-Jea Ko. Improving the usability
of remote eye gaze tracking for human-device interaction. IEEE
Transactions on Consumer Electronics, 60(3):493–498, 2014.
[123] Jaegul Choo and Shixia Liu. Visual analytics for explainable deep
learning. IEEE Computer Graphics and Applications, 38(4):84–92,
2018.
[124] William S. Cleveland and Robert McGill. An experiment in graphical
perception. International Journal of Man-Machine Studies, 25(5):491–
501, 1986.
[125] Jean Clottes. Chauvet Cave: The Art of Earliest Times. University of
Utah Press, 2003.
[126] Moreno I. Coco. The statistical challenge of scan-path analysis. In
Proceedings of the 2nd conference on Human System Interactions, HSI,
pages 369–372, 2009.
[127] Christopher Collins, Gerald Penn, and Sheelagh Carpendale.
Bubble sets: Revealing set relations with isocontours over existing
visualizations. IEEE Transactions on Visualization and Computer
Graphics, 15(6):1009–1016, 2009.
[128] Cristina Conati, Christina Merten, Saleema Amershi, and Kasia
Muldner. Using eye-tracking data for high-level user modeling in
adaptive interfaces. In Proceedings of the Twenty-Second AAAI
Conference on Artificial Intelligence, pages 1614–1617. AAAI Press,
2007.
[129] Gregorio Convertino, Jian Chen, Beth Yost, Young-Sam Ryu, and
Chris North. Exploring context switching and cognition in dual-view
288 References
[139] Aaron Crug and Rochel Gelman. Counting is a piece of cake: Yet
another counting task. In Laura A. Carlson, Christoph Hölscher, and
Thomas F. Shipley, editors, Proceedings of the 33th Annual Meeting
of the Cognitive Science Society, CogSci. cognitivesciencesociety.org,
2011.
[140] James E. Cutting. How the eye measures reality and virtual reality.
Behavior Research Methods, Instruments, and Computers, 29:27–36,
1997.
[141] Filip Dabek and Jesus J. Caban. A grammar-based approach for
modeling user interactions and generating suggestions during the
data exploration process. IEEE Transactions on Visualization and
Computer Graphics, 23(1):41–50, 2017.
[142] Robertas Damasevicius, Jevgenijus Toldinas, Algimantas
Venckauskas, Sarunas Grigaliunas, Nerijus Morkevicius, and Vaidas
Jukavicius. Visual analytics for cyber security domain: State-of-the-art
and challenges. In Robertas Damasevicius and Giedre Vasiljeviene,
editors, Proceedings of 25th International Conference on Information
and Software Technologies, ICIST, volume 1078 of Communications
in Computer and Information Science, pages 256–270. Springer, 2019.
[143] Subhrajit Das, Tom McEwan, and Donna Douglas. Using eye-tracking
to evaluate label alignment in online forms. In Agneta Gulz, Charlotte
Magnusson, Lone Malmborg, Håkan Eftring, Bodil Jönsson, and
Konrad Tollmar, editors, Proceedings of the 5th Nordic Conference
on Human-Computer Interaction, volume 358 of ACM International
Conference Proceeding Series, pages 451–454. ACM, 2008.
[144] Alan Davies, Markel Vigo, Simon Harper, and Caroline Jay. The
visualisation of eye-tracking scanpaths: what can they tell us about
how clinicians view electrocardiograms? In Michael Burch, Lewis L.
Chuang, and Andrew T. Duchowski, editors, Proceedings of IEEE
Workshop on Eye Tracking and Visualization, ETVIS, pages 79–83.
IEEE Computer Society, 2016.
[145] Marcelo de Paiva Guimarães, Diego Roberto Colombo
Dias, José Hamilton Mota, Bruno Barberi Gnecco, Vinicius
Humberto Serapilha Durelli, and Luis Carlos Trevelin. Immersive
and interactive virtual reality applications based on 3d web browsers.
Multimedia Tools and Applications, 77(1):347–361, 2018.
[146] Thomas A. DeFanti, Maxine D. Brown, and Bruce H. McCormick.
Visualization: Expanding scientific and engineering research
opportunities. Computer, 22(8):12–25, 1989.
290 References
[147] Philippe Dessus, Olivier Cosnefroy, and Vanda Luengo. "keep your
eyes on ’em all!": A mobile eye-tracking analysis of teachers’
sensitivity to students. In Katrien Verbert, Mike Sharples, and Tomaz
Klobucar, editors, Proceedings of the 11th European Conference on
Technology Enhanced Learning, Adaptive and Adaptable Learning ,
EC-TEL, volume 9891 of Lecture Notes in Computer Science, pages
72–84. Springer, 2016.
[148] Richard Dewhurst, Marcus Nyström, Halszka Jarodzka, Tom
Foulsham, Roger Johansson, and Kenneth Holmqvist. It depends on
how you look at it: Scanpath comparison in multiple dimensions with
MultiMatch, a vector-based approach. Behavior Research Methods,
44:1079 – 1100, 2012.
[149] Stephan Diehl. Software Visualization - Visualizing the Structure,
Behaviour, and Evolution of Software. Springer, 2007.
[150] Stephan Diehl, Fabian Beck, and Michael Burch. Uncovering strengths
and weaknesses of radial visualizations—an empirical approach. IEEE
Transactions on Visualization and Computer Graphics, 16(6):935–
942, 2010.
[151] Stephan Diehl and Carsten Görg. Graphs, they are changing. In
Stephen G. Kobourov and Michael T. Goodrich, editors, Proceedings
of the 10th International Symposium on Graph Drawing, GD, volume
2528 of Lecture Notes in Computer Science, pages 23–30. Springer,
2002.
[152] Evanthia Dimara and Charles Perin. What is interaction for data
visualization? IEEE Transactions on Visualization and Computer
Graphics, 26(1):119–129, 2020.
[153] Dimitra Dimitrakopoulou, Evanthia Faliagka, and Maria Rigou. An
eye-tracking study of web search interaction design patterns. In
Proceedings of the 20th Pan-Hellenic Conference on Informatics,
page 22. ACM, 2016.
[154] Alan J. Dix. Designing for appropriation. In Thomas C. Ormerod and
Corina Sas, editors, Proceedings of the 21st British HCI Group Annual
Conference on HCI 2007: HCI...but not as we know it - Volume 2, BCS
HCI, pages 27–30. BCS, 2007.
[155] Timothy D. Dixon, Jian Li, Jan M. Noyes, Tom Troscianko, Stavri G.
Nikolov, John J. Lewis, Eduardo Fernández Canga, David R. Bull, and
Cedric Nishan Canagarajah. Scanpath analysis of fused multi-sensor
images with luminance change: A pilot study. In Proceedings of 9th
References 291
topographic maps, optical, and radar data in the northern nile delta,
egypt. Remote Sensing, 11(24):3039, 2019.
[171] Geoffrey P. Ellis and Alan J. Dix. An explorative analysis of user
evaluation studies in information visualisation. In Enrico Bertini,
Catherine Plaisant, and Giuseppe Santucci, editors, Proceedings of the
AVI Workshop on BEyond time and errors: novel evaluation methods
for information visualization, BELIV, pages 1–7. ACM Press, 2006.
[172] Niklas Elmqvist, Pierre Dragicevic, and Jean-Daniel Fekete. Rolling
the dice: Multidimensional visual exploration using scatterplot
matrix navigation. IEEE Transactions on Visualization and Computer
Graphics, 14(6):1539–1148, 2008.
[173] Alex Endert, Patrick Fiaux, and Chris North. Semantic interaction for
sensemaking: Inferring analytical reasoning for model steering. IEEE
Transactions on Visualization and Computer Graphics, 18(12):2879–
2888, 2012.
[174] Alex Endert, Mahmud Shahriar Hossain, Naren Ramakrishnan, Chris
North, Patrick Fiaux, and Christopher Andrews. The human is the loop:
new directions for visual analytics. Journal of Intelligent Information
Systems, 43(3):411–435, 2014.
[175] Sukru Eraslan, Yeliz Yesilada, and Simon Harper. Scanpath trend
analysis on web pages: Clustering eye tracking scanpaths. ACM
Transactions on the Web, 10(4):20:1–20:35, 2016.
[176] Mateus Espadoto, Rafael Messias Martins, Andreas Kerren, Nina S. T.
Hirata, and Alexandru C. Telea. Toward a quantitative survey of
dimension reduction techniques. IEEE Transactions on Visualization
and Computer Graphics, 27(3):2153–2173, 2021.
[177] Augusto Esteves, Eduardo Velloso, Andreas Bulling, and Hans
Gellersen. Orbits: Gaze interaction for smart watches using smooth
pursuit eye movements. In Celine Latulipe, Bjoern Hartmann, and Tovi
Grossman, editors, Proceedings of the 28th Annual ACM Symposium
on User Interface Software & Technology, UIST, pages 457–466.
ACM, 2015.
[178] Ronak Etemadpour, Bettina Olk, and Lars Linsen. Eye-tracking
investigation during visual analysis of projected multidimensional data
with 2d scatterplots. In Robert S. Laramee, Andreas Kerren, and
José Braz, editors, Proceedings of the 5th International Conference
on Information Visualization Theory and Applications, IVAPP, pages
233–246. SciTePress, 2014.
294 References
[206] Robert B. Goldstein, Russell L. Woods, and Eli Peli. Where people
look when watching movies: Do all viewers look at the same place?
Computers in Biology and Medicine, 37(7):957–964, 2007.
[207] David Gotz, Shun Sun, and Nan Cao. Adaptive contextualization:
Combating bias during high-dimensional visualization and data
selection. In Jeffrey Nichols, Jalal Mahmud, John O’Donovan, Cristina
Conati, and Massimo Zancanaro, editors, Proceedings of the 21st
International Conference on Intelligent User Interfaces, IUI, pages
85–95. ACM, 2016.
[208] Natalia Grabar, Emmanuel Farce, and Laurent Sparrow. Study of
readability of health documents with eye-tracking methods. In Pascale
Sébillot and Vincent Claveau, editors, Actes de la Conférence, CORIA-
TALN-RJC, pages 3–18. ATALA, 2018.
[209] Thomas Grindinger, Andrew T. Duchowski, and Michael W. Sawyer.
Group-wise similarity and classification of aggregate scanpaths. In
Carlos Hitoshi Morimoto, Howell O. Istance, Aulikki Hyrskykari, and
Qiang Ji, editors, Proceedings of the Symposium on Eye-Tracking
Research & Applications, ETRA, pages 101–104. ACM, 2010.
[210] Martin Groen and Jan Noyes. Using eye tracking to evaluate usability
of user interfaces: Is it warranted? In Frédéric Vanderhaegen, editor,
Proceedings of 11th IFAC/IFIP/IFORS/IEA Symposium on Analysis,
Design, and Evaluation of Human-Machine Systems, pages 489–493.
International Federation of Automatic Control, 2010.
[211] Yi Gu, Chaoli Wang, Robert Bixler, and Sidney K. D’Mello. Etgraph:
A graph-based approach for visual analytics of eye-tracking data.
Computers & Graphics, 62:1–14, 2017.
[212] Drew T. Guarnera. Enhancing eye tracking of source code: A
specialized fixation filter for source code. In Proceedings of the IEEE
International Conference on Software Maintenance and Evolution,
ICSME, pages 615–618. IEEE, 2019.
[213] Egon Guba, Willavene Wolf, Sybil de Groot, Manfred Knemeyer,
Ralph van Atta, and Larry Light. Eye movements and tv viewing in
children. AV Communication Review, 12(4):386 – 401, 1964.
[214] Wei Guo and Shiwei Cheng. An approach to reading assistance with
eye tracking data and text features. In Proceedings of Adjunct of
the 2019 International Conference on Multimodal Interaction, ICMI,
pages 7:1–7:7. ACM, 2019.
[215] Jacek Gwizdka and Michael J. Cole. Does interactive search results
overview help?: an eye tracking study. In Wendy E. Mackay,
298 References
[224] Julian Heinrich and Daniel Weiskopf. State of the art of parallel
coordinates. In Mateu Sbert and László Szirmay-Kalos, editors,
Proceedings of 34th Annual Conference of the European Association
for Computer Graphics, Eurographics - State of the Art Reports, pages
95–116. Eurographics Association, 2013.
[225] Helene Hembrooke, Matthew K. Feusner, and Geri Gay. Averaging
scan patterns and what they can tell us. In Kari-Jouko Räihä and
Andrew T. Duchowski, editors, Proceedings of the Eye Tracking
Research & Application Symposium, ETRA, page 41. ACM, 2006.
[226] Bradley M. Hemminger, Trish Losi, and Anne Bauers. Survey of
bioinformatics programs in the united states. Journal of the American
Society for Information Science and Technology (JASIST), 56(5):529–
537, 2005.
[227] Nathalie Henry, Jean-Daniel Fekete, and Michael J. McGuffin.
Nodetrix: a hybrid visualization of social networks. IEEE Transactions
on Visualization and Computer Graphics, 13(6):1302–1309, 2007.
[228] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based
computational model. In Hugo Larochelle, Marc’Aurelio Ranzato,
Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors,
Proceedings of Advances in Neural Information Processing Systems
33: Annual Conference on Neural Information Processing Systems,
NeurIPS, 2020.
[229] Marcel Hlawatsch, Michael Burch, and Daniel Weiskopf. Visual
adjacency lists for dynamic graphs. IEEE Transactions on
Visualization and Computer Graphics, 20(11):1590–1603, 2014.
[230] Marcel Hlawatsch, Filip Sadlo, Michael Burch, and Daniel Weiskopf.
Scale-stack bar charts. Computer Graphics Forum, 32(3):181–190,
2013.
[231] Hsin Yang Ho, I-Cheng Yeh, Yu-Chi Lai, Wen-Chieh Lin, and Fu-Yin
Cherng. Evaluating 2d flow visualization using eye tracking. Computer
Graphics Forum, 34(3):501–510, 2015.
[232] Benjamin Höferlin. Scalable Visual Analytics for Video Surveillance.
PhD thesis, University of Stuttgart, 2014.
[233] Benjamin Höferlin, Markus Höferlin, Gunther Heidemann, and Daniel
Weiskopf. Scalable video visual analytics. Information Visualization,
14(1):10–26, 2015.
[234] Benjamin Höferlin, Rudolf Netzel, Markus Höferlin, Daniel Weiskopf,
and Gunther Heidemann. Inter-active learning of ad-hoc classifiers for
video visual analytics. In Proceedings of 7th IEEE Conference on
300 References
[354] Corine S. Meppelink and Nadine Bol. Exploring the role of health
literacy on attention to and recall of text-illustrated health information:
An eye-tracking study. Computers in Human Behavior, 48:87–93,
2015.
[355] Slavko Milekic. Using eye- and gaze-tracking to interact with a visual
display. In Stuart Dunn, Jonathan P. Bowen, and Kia Ng, editors,
Proceedings of Electronic Visualisation and the Arts, EVA, Workshops
in Computing. BCS, 2012.
[356] Kazuo Misue, Peter Eades, Wei Lai, and Kozo Sugiyama. Layout
adjustment and the mental map. Journal of Visual Languages and
Computing, 6(2):183–210, 1995.
[357] Parag K. Mital, Tim J Smith, Robin L. Hill, and John M. Henderson.
Clustering of gaze during dynamic scene viewing is predicted by
motion. Cognitive Computation, 3(1):5–24, 2011.
[358] Nadine Marie Moacdieh and Nadine Sarter. Using eye tracking to
detect the effects of clutter on visual search in real time. IEEE
Transactions on Human-Machine Systems, 47(6):896–902, 2017.
[359] Kenneth Moreland. A survey of visualization pipelines. IEEE
Transactions on Visualization and Computer Graphics, 19(3):367–
378, 2013.
[360] Carlos Hitoshi Morimoto and Marcio R. M. Mimica. Eye gaze tracking
techniques for interactive applications. Computer Vision and Image
Understanding, 98(1):4–24, 2005.
[361] Rajeev Motwani and Prabhakar Raghavan. Randomized Algorithms.
Cambridge University Press, 1995.
[362] Haris Mumtaz, Shahid Latif, Fabian Beck, and Daniel Weiskopf.
Exploranative code quality documents. IEEE Transactions on
Visualization and Computer Graphics, 26(1):1129–1139, 2020.
[363] Tanja Munz, Michael Burch, Toon van Benthem, Yoeri Poels, Fabian
Beck, and Daniel Weiskopf. Overlap-free drawing of generalized
pythagoras trees for hierarchy visualization. In Proceedings of 30th
IEEE Visualization Conference, IEEE VIS - Short Papers, pages
251–255. IEEE, 2019.
[364] Tamara Munzner. Process and Pitfalls in Writing Information
Visualization Research Papers, pages 134–153. Springer Berlin
Heidelberg, Berlin, Heidelberg, 2008.
[365] Minoru Nakayama and Yuko Hayashi. Estimation of viewer’s response
for contextual understanding of tasks using features of eye-movements.
In Carlos Hitoshi Morimoto, Howell O. Istance, Aulikki Hyrskykari,
References 313
[393] Fabio Pellacini, Lori Lorigo, and Geri Gay. Visualizing paths in
context. In Technical Report, Computer Science, 2006.
[394] Thies Pfeiffer. Measuring and visualizing attention in space with 3d
attention volumes. In Carlos Hitoshi Morimoto, Howell O. Istance,
Stephen N. Spencer, Jeffrey B. Mulligan, and Pernilla Qvarfordt,
editors, Proceedings of the Symposium on Eye-Tracking Research and
Applications, ETRA, pages 29–36. ACM, 2012.
[395] Thies Pfeiffer, Patrick Renner, and Nadine Pfeiffer-Leßmann.
Eyesee3d 2.0: model-based real-time analysis of mobile eye-tracking
in static and dynamic three-dimensional scenes. In Pernilla Qvarfordt
and Dan Witzner Hansen, editors, Proceedings of the Ninth Biennial
ACM Symposium on Eye Tracking Research & Applications, ETRA,
pages 189–196. ACM, 2016.
[396] William A. Pike, John T. Stasko, Remco Chang, and Theresa A.
O’Connell. The science of interaction. Information Visualization,
8(4):263–274, 2009.
[397] Catherine Plaisant. The challenge of information visualization
evaluation. In Maria Francesca Costabile, editor, Proceedings of the
working conference on Advanced visual interfaces, AVI, pages 109–
116. ACM Press, 2004.
[398] Catherine Plaisant, Jesse Grosjean, and Benjamin B. Bederson.
Spacetree: Supporting exploration in large node link tree, design
evolution and empirical evaluation. In Pak Chung Wong and Keith
Andrews, editors, Proceedings of IEEE Symposium on Information
Visualization (InfoVis), pages 57–64. IEEE Computer Society, 2002.
[399] Christopher Plaue, Todd Miller, and John T. Stasko. Is a picture worth
a thousand words? an evaluation of information awareness displays.
In Wolfgang Heidrich and Ravin Balakrishnan, editors, Proceedings of
the Graphics Interface Conference, volume 62 of ACM International
Conference Proceeding Series, pages 117–126. Canadian Human-
Computer Communications Society, 2004.
[400] Mathias Pohl, Markus Schmitt, and Stephan Diehl. Comparing
the readability of graph layouts using eyetracking and task-
oriented analysis. In Proceedings of the Eurographics Workshop on
Computational Aesthetics, pages 49–56, 2009.
[401] Alex Poole and Linden J. Ball. Eye tracking in human-computer
interaction and usability research: Current status and future prospects.
In Idea Group Inc., ch. E, pages 211–219, 2006.
References 317
[473] Oleg Spakov and Darius Miniotas. Visualization of eye gaze data using
heat maps. Electronics and Electrical Engineering, 74:55–58, 2007.
[474] Oleg Spakov and Kari-Jouko Räihä. KiEV: a tool for visualization
of reading and writing processes in translation of text. In Kari-Jouko
Räihä and Andrew T. Duchowski, editors, Proceedings of the Eye
Tracking Research & Application Symposium, ETRA, pages 107–110.
ACM, 2008.
[475] Robert Spence. Rapid, serial and visual: a presentation technique with
potential. Information Visualization, 1(1):13–19, 2002.
[476] Robert Spence. Information Visualization: Design for Interaction.
Pearson/Prentice Hall, 2 edition, 2007.
[477] Robert Spence and Mark Witkowski. Rapid Serial Visual Presentation
- Design for Cognition. Springer Briefs in Computer Science. Springer,
2013.
[478] Bernd Carsten Stahl and David Wright. Ethics and privacy in AI and
big data: Implementing responsible research and innovation. IEEE
Security & Privacy, 16(3):26–33, 2018.
[479] John T. Stasko. Tango: A framework and system for algorithm
animation. Computer, 23(9):27–39, 1990.
[480] John T. Stasko. Animating algorithms with XTANGO. SIGACT News,
23(2):67–71, 1992.
[481] Ben Steichen, Giuseppe Carenini, and Cristina Conati. User-adaptive
information visualization: using eye gaze data to infer visualization
tasks and user cognitive abilities. In Jihie Kim, Jeffrey Nichols,
and Pedro A. Szekely, editors, Proceedings of 18th International
Conference on Intelligent User Interfaces, IUI, pages 317–328. ACM,
2013.
[482] Ben Steichen, Oliver Schmid, Cristina Conati, and Giuseppe Carenini.
Seeing how you’re looking - using real-time eye gaze data for user-
adaptive visualization. In Shlomo Berkovsky, Eelco Herder, Pasquale
Lops, and Olga C. Santos, editors, Proceedings of the 21st Conference
on User Modeling, Adaptation, and Personalization, volume 997.
CEUR-WS.org, 2013.
[483] Sophie Stellmach, Lennart E. Nacke, and Raimund Dachselt. 3d
attentional maps: aggregated gaze visualizations in three-dimensional
virtual environments. In Giuseppe Santucci, editor, Proceedings of the
International Conference on Advanced Visual Interfaces, AVI, pages
345–348. ACM Press, 2010.
326 References
[494] Soon Tee Teoh. A study on multiple views for tree visualization. In
Proceedings of the Conference on Visualization and Data Analysis,
volume 6495, 2007.
[495] James J. Thomas and Kristin A. Cook. Illuminating the Path - The
Research and Development Agenda for Visual Analytics. National
Visualization and Analytics Ctr, 2005.
[496] Christian Tominski. Interaction for Visualization. Synthesis Lectures
on Visualization. Morgan & Claypool Publishers, 2015.
[497] Warren S. Torgerson. Multidimensional scaling I: Theory and method.
Psychometrika, 17:401–419, 1952.
[498] Melanie Tory, M. Stella Atkins, Arthur E. Kirkpatrick, Marios
Nicolaou, and Guang-Zhong Yang. Eyegaze analysis of displays
with combined 2d and 3d views. In Proceedings of the 16th IEEE
Visualization Conference, IEEE, pages 519–526. IEEE Computer
Society, 2005.
[499] Michael Traoré and Christophe Hurter. Exploratory study with eye
tracking devices to build interactive systems for air traffic controllers.
In Guy A. Boy, editor, Proceedings of the International Conference on
Human-Computer Interaction in Aerospace, HCI-Aero, pages 6:1–6:9.
ACM, 2016.
[500] Anne Treisman. Preattentive processing in vision. Computer Vision,
Graphics, and Image Processing, 31(2):156–177, 1985.
[501] Sandra Trösterer, Alexander Meschtscherjakov, David Wilfinger, and
Manfred Tscheligi. Eye tracking in the car: Challenges in a dual-
task scenario on a test track. In Linda Ng Boyle, Gary E. Burnett,
Peter Fröhlich, Shamsi T. Iqbal, Erika Miller, and Yuqing Wu,
editors, Adjunct Proceedings of the 6th International Conference on
Automotive User Interfaces and Interactive Vehicular Applications,
pages 12:1–12:6. ACM, 2014.
[502] Hoi Ying Tsang, Melanie Tory, and Colin Swindells. eSeeTrack
- visualizing sequential fixation patterns. IEEE Transactions on
Visualization and Computer Graphics, 16(6):953–962, 2010.
[503] Edward Rolf Tufte. The visual display of quantitative information.
Graphics Press, 1992.
[504] Vilius Turenko, Simonas Baltulionis, Mindaugas Vasiljevas, and
Robertas Damasevicius. Analysing program source code reading skills
with eye tracking technology. In Robertas Damasevicius, Tomas
Krilavicius, Audrius Lopata, Dawid Polap, and Marcin Wozniak,
editors, Proceedings of the International Conference on Information
328 References
[535] Tohru Yagi, Yoshiaki Kuno, Kazuo Koga, and Toshiharu Mukai.
Drifting and blinking compensation in electro-oculography (EOG)
eye-gaze interface. In Proceedings of the IEEE International
Conference on Systems, Man and Cybernetics, pages 3222–3226.
IEEE, 2006.
[536] Mehmet Adil Yalçin, Niklas Elmqvist, and Benjamin B. Bederson.
Raising the bars: Evaluating treemaps vs. wrapped bars for
dense visualization of sorted numeric data. In Elmar Eisemann
and Scott Bateman, editors, Proceedings of the 43rd Graphics
Interface Conference, pages 41–49. Canadian Human-Computer
Communications Society / ACM, 2017.
[537] Yasuko Yamagishi, Hiroshi Yanaka, Katsuhiko Suzuki, Seiji Tsuboi,
Takehi Isse, Masayuki Obayashi, Hajimu Tamura, and Hiromichi
Nagao. Visualization of geoscience data on google earth: Development
of a data converter system for seismic tomographic models. Computers
& Geosciences, 36(3):373–382, 2010.
[538] Chia-Kai Yang. Identifying reading patterns with eye-tracking visual
analytics. In Andreas Bulling, Anke Huckauf, Eakta Jain, Ralph
Radach, and Daniel Weiskopf, editors, Proceedings of the Symposium
on Eye Tracking Research and Applications, ETRA Adjunct, pages
10:1–10:3. ACM, 2020.
[539] Alfred L. Yarbus. Eye Movements and Vision. Springer, 1967.
[540] Lora Yekhshatyan and John D. Lee. Changes in the correlation
between eye and steering movements indicate driver distraction. IEEE
Transactions on Intelligent Transportation Systems, 14(1):136–145,
2013.
[541] Chi-Hsien Yen, Aditya G. Parameswaran, and Wai-Tat Fu. An
exploratory user study of visual causality analysis. Computer Graphics
Forum, 38(3):173–184, 2019.
[542] Leelakrishna Yenigalla, Vinayak Sinha, Bonita Sharif, and Martha E.
Crosby. How novices read source code in introductory courses
on programming: An eye-tracking experiment. In Dylan D.
Schmorrow and Cali M. Fidopiastis, editors, Proceedings of 10th
International Conference on Foundations of Augmented Cognition:
Neuroergonomics and Operational Neuroscience, AC, volume 9744 of
Lecture Notes in Computer Science, pages 120–131. Springer, 2016.
[543] Yeliz Yesilada, Caroline Jay, Robert Stevens, and Simon Harper.
Validating the use and role of visual elements of web pages in
navigation with an eye-tracking study. In Jinpeng Huai, Robin Chen,
332 References
335
336 Index
Visual metaphor 26 W
Visual pattern 21 Walk-and-interact 216
Visual search pattern 212 Walk-then-interact 216
Visual stimulus 149, 199 Wearable eye tracker 132, 189, 191
Visual system 177 Web-based environment 38
Visual task solution strategy 213 Web-base visualization 66
Visual transformation 20 Wedge chart 34
Visual variable 19, 22 Weighted browsing 51, 121
Visualization 35 Welch’s test 162
Visualization framework 38 William Playfair 32
Visualization library 38 Within-subjects design 129, 147
Visualization pipeline 54 Word cloud 46
Visualization technique 254 World wide web 38
Voice 59 Wrapped bar 169
Voice recognition 194
Volumetric data 47 Z
Voronoi region 206 Zooming 31
About the Author
347