Interaction in the Visualization of Multivariate Networks

Michael Wybrow, Björn Zimmer, Niklas Elmqvist, Jean-Daniel Fekete,

Tatiana von Landesberger, Jarke van Wijk

To cite this version:

Michael Wybrow, Björn Zimmer, Niklas Elmqvist, Jean-Daniel Fekete, Tatiana von Landesberger, et
al.. Interaction in the Visualization of Multivariate Networks. Kerren, Andreas, Purchase, Helen,
Ward, Matthew O. Multivariate Network Visualization, 8380, Springer, pp.XV, 237, 2014, Lecture
Notes in Computer Science, 978-3-319-06792-6. �hal-00974335v1�

HAL Id: hal-00974335

https://hal.inria.fr/hal-00974335v1
Submitted on 6 Apr 2014 (v1), last revised 13 Apr 2014 (v2)

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Interaction in the Visualization
of Multivariate Networks

Michael Wybrow1 , Björn Zimmer2 Niklas Elmqvist3 , Jean-Daniel Fekete4 ,

Tatiana von Landesberger5 , and Jarke J. van Wijk6
1
Monash University, Melbourne, Australia
[email protected]
2
Linnaeus University, Växjö, Sweden
[email protected]
3
Purdue University, West Lafayette, IN, USA
[email protected]
4
INRIA, France
[email protected]
5
TU Darmstadt, Darmstadt, Germany
[email protected]
6
Eindhoven University of Technology, Eindhoven, The Netherlands
[email protected]

DRAFT: October 18, 2013

1 Introduction

The overall aim of visualization is to obtain insight into large amounts of data.
Detection of patterns as well as outliers are typical examples. For networks, such
patterns can be number and position of cliques; for multivariate data this can be
the correlation between attributes. The major challenge of multivariate network
visualization is to understand the interplay between properties of the network
and its associated data, for instance to see if the formation of cliques can be
understood from attributes of nodes.
Producing useful and informative visualizations for multivariate networks is
a complex and challenging task. One of the main issues is scalability both with
respect to the graph size as well as to the number and variety of variables. It
is very difficult to statically display large data sets in general, including mul-
tivariate data and networks. Occasionally it is possible to nicely encode small
multivariate data sets completely in custom static visualizations, such as with
Minard’s seminal “Napoleon’s March to Moscow” visualization [49], but this is
rare.
In practice, even moderate-sized networks pose problems when displaying
them without causing overlaps and loss of information, let alone augmenting
these with additional variables. Moreover, a person can usually only comprehend
a small subset of the information space at a time. It is therefore important to
reduce the relevant information displayed at any point to a manageable amount
2

in order to facilitate understanding of the main data characteristics. Thus, as

the data size and complexity (i.e., the combination of dimensions and network
complexity) increases, there is a need to efficiently navigate through the data
and to enable discovery and communication of the data.
Interaction is a vital component in the visualization of multivariate networks.
By allowing people to browse data sets with interactions like panning and zoom-
ing, it enables much more information to be seen and explored than would oth-
erwise be possible with static visualization. Overview-based interactions afford
the user the ability to understand a complete picture of the data or informa-
tion landscape and to decide where to direct her attention. Through search and
filtering, interaction can reduce cognitive effort on users by allowing them to
locate, focus on and understand subsets of the data in isolation. Pivoting and
other navigational interactions at both the view- and data-level allow people to
identify and then to transition between areas of interest.
While there are methods for interacting with graphs and dimensions sep-
arately, the combination of both needs special attention. The challenge is to
clearly visualize multiple sets of individual dimensions as well as to offer a useful
visual overview of data, and allow transitions between these to be easily under-
stood. Moreover, we need to find ways to support users in navigating through
the complex data space (graphs x dimensions) without “getting lost” without
an overburden of interaction actions, as this might me frustrating for the user.
TODO: Refer back to Domain Applications chapter by mention-
ing examples of a successful or interesting systems or visualization
approaches from each of our domains, with specific focus on the in-
teraction techniques employed.
This chapter aims to illustrate the importance of interaction for the visualiza-
tion of multivariate networks. It will discuss the design space, examine existing
approaches, offer some guidance for designing interactions and describe open
problems in the area. It is aimed at a reader who is intending to visualize net-
works with multivariate data. They may be planning to evaluate and select some
existing approaches or systems and adapt these to their needs, or they may be
thinking about designing a custom visualization tailored to the needs of their
data and audience. Rather than just a survey of the field, this chapter should be
considered a guide to interaction for networks with multivariate data; explaining
what the problems are, what is possible, what has been done before, what might
be done in future.
The rest of the chapter is organized in five further sections. The next section
discusses the design space and requirements of working with large multivariate
data sets, including difficulties in navigating networks and dimensions. Section 3
classifies relevant interaction techniques on the basis of the stages in the standard
Information Visualization reference model. Section 4 gives examples of the in-
teractive aspects of multivariate graph visualization systems. Section 5 presents
recommendations and guidelines for designing novel interaction approaches, in-
cluding adaptation of existing interaction design principles for use in this setting.
 3

Data
Visual
Form
U
Raw Data Visual
Views
Data Tables Structures

Data Visual View User/Task

Transformations Mappings Transformations

Human Interaction

Fig. 1. The information visualization reference model.

Finally, Section 6 puts forward a vision of the challenges and goals as we see them
within the field of multivariate graph visualization.

2 Background

Interaction is a vital ingredient of information visualization, and has been heavily

studied. In this section, we do not aim to explain in general how interaction works
in visualization, as this is very well addressed by excellent books such as [53] and
a large number of articles. Also, we acknowledge that data exploration encom-
passes much more than just direct interaction with graphical representations,
and includes aspects like navigation support, knowledge capture, and collabo-
rative visualization. This area is studied in visual analytics; for an overview see
Pike et al. [47].
Furthermore, for this chapter, we mostly consider interaction for standard
point-and-click and keyboard interfaces on desktop computers. While multi-
touch tablets are commonplace and we are seeing increasing availability of large
touch-based tables and displays, there has been relatively little work designing or
evaluating interaction techniques for working with large networks or multivari-
ate data on these. This is also the case with other new technology now becoming
available to consumers such as 3D displays, contactless input devices, and multi-
monitor displays. We discuss this as a key ongoing challenge in Section 6.
Data exploration often involves a top down approach, as strongly summarized
in the visual information seeking mantra of Ben Shneiderman [51]: “overview
first, zoom and filter, details on demand”. Both for network and multivariate vi-
sualization, many systems and techniques aim to satisfy this pattern. But also,
in practice a bottom up approach is used. For instance, in social network visu-
alization a certain person can be the starting point for further exploration [27];
in multivariate visualization one can start from one particular item and explore
which are similar. As both approaches are valuable, an ideal system should sup-
port both.
4

To describe the multiple kinds of interactions used for the visualization of

multivariate networks in more detail, we use the Information Visualization Ref-
erence Model [10] (see Figure 1), which breaks down the visualization process
into four stages: raw data or source data, data tables, visual structures or visual
abstractions, and views. To display the raw data, several transformations have to
be applied: the raw data is transformed into data tables through data transfor-
mations, the data tables into visual structures through visual mappings, and the
final rendering transforms the visual structure into a image on a view. All these
transformations are performed using a multitude of specific parameters, and in-
teraction then can be defined at the system level as the change of transformation
parameters controlled by the user with immediate feedback to the user.
This generic model applies both to network and multivariate visualization,
and many interaction techniques specifically tailored to the properties of these
data types have been developed. In the next section we enumerate the most
relevant of these, categorized along the stages of the reference model. However,
far less techniques have been developed that specifically aim at interaction with
combinations of network and multivariate data. The challenge here is to offer
a simple but powerful set of interaction techniques that allows users to explore
such combinations with minimal cognitive overload. On one hand, this should
be doable, since many tasks and operations are similar at a high level; but on
the other hand, standard representations of networks and multivariate data do
vary largely, and also the more powerful and customized interaction methods for
dealing with these data types differ strongly.
These effects can be observed for all stages of the reference model. At first
sight, network data and multivariate data seem fundamentally different. How-
ever, topological aspects of network data can be nicely captured as multivariate
data, simply by calculating topological metrics of nodes and edges. Also, mul-
tivariate data can be considered as networks, for instance by introducing edges
between nodes that are similar, as pursued by Liu et al. [38]. Having said this,
multivariate network visualization cannot usually be reduced to purely multi-
variate or network visualization. In fact, the combination makes analysis of mul-
tivariate networks a real challenge since discovery of an underlying phenomenon
in the data can require a detailed understanding of the network topology to-
gether with the multivariate attributes, e.g., if variables represent snapshots of
a flow dictated by the topology. One consequence for interaction is that users
should be enabled to obtain such associated data on request. Filtering of data is a
standard operation. For multivariate data this typically involves selection based
on ranges of attributes, for networks for instance the distance from a selected
set of nodes can be used.
Concerning the visual representation, network data and multivariate data
can be shown separately or be combined. The use of multiple views on data
is standard in visualization, and by interacting through linking and brushing,
information from different views can be associated. Interaction is crucial here,
but also, as both types of data are shown separately, fusion of information is
often hard. One way to provide a combined view is to use a network based
 5

approach, where nodes and edges are embellished with iconic representations
of values or attributes. This limits the use of standard interaction methods for
multivariate data, for instance to select two ranges for attributes by sweeping out
a rectangle in a scatterplot. Another way to combine data in one view is to use
a multivariate data based approach, for instance by superimposing edges on top
of a scatterplot. Now, standard interaction methods for multivariate data can
be used, as positions of nodes encode attribute values, but also, some network
interaction techniques that imply changes in the layout cannot be used anymore.
The standard approach in the view transformation stage is to provide options
for zooming and panning. On the image level, this is straightforward, however,
when using multiple views where the spatial dimensions have different meanings,
this can be hard to deal with in a natural way.
These examples show there are basically two approaches to interacting with
multivariate networks. One approach is to stick to conventional representations
and dedicated interaction methods, another, more challenging but also poten-
tially more rewarding approach is to aim for tight integration, both with respect
to representation and interaction, to facilitate the understanding of the relation
between network and multivariate. In the following sections these approaches
are explored in more detail.

3 Classification of Interactions

We use the information visualization reference model, originally presented in [10],

as the basis for our classification of interaction techniques (see Figure 1). We
classify interaction techniques based on the level of this pipeline they affect.
The match is not always perfect, as some techniques address multiple levels
simultaneously. Where possible, we make use of standard terminology and jargon
from the information visualization community in order to simplify access to
related work.
Notably, our classification presents the pipeline stages in the reverse order
to [10]: we describe interactions at the view level first for pedagogical clarity,
since these are simpler, easier to understand, and are sometimes extended or
utilized by interactions in the remaining stages of the pipeline.
Many of the generic interaction techniques are applicable both to standard
networks as well as multivariate data, and basic examples are given. As discussed
in Section 2, there are many possible graph representations, the choice of which
can limit the applicability of interaction techniques since these may be dependant
on specific aspects of the chosen graph representation. Examples of complete
systems utilizing a mix of interaction techniques to deal simultaneously with a
combination multivariate data and networks are described in Section 4.
This classification is a revised and expanded version of a similar classification
of interaction techniques for network visualization appearing in [63]. Note, this
is certainly not the only way to define and categorize interaction. For instance,
Yi et al. advocate for a taxonomy based on user intent, and they distinguish
6

Select, Explore, Reconfigure, Encode, Abstract/Elaborate, Filter, and Connect

as main categories [67].

3.1 View Level Interactions

The view-level interactions are mostly related to highlighting of interesting ob-

jects, navigation through the data set, and using Magic Lenses to augment the
visualized information.

Highlighting Highlighting transiently changes the visible rendition of items at

the view level, not at the visual encoding level. Although it can be practically
implemented with support at the visual structures level, this is not required so
we conceptually consider it a view-level interaction.
Interactions such as search or mouse hovering may lead to highlighting of
objects such as search results or linked content.

Hovering. Hovering is used in multivariate visualizations such as InfoZoom [56]

that display large data tables with a smart aggregation mechanism. Rows are
items, columns are attributes, and values are in cells. When the mouse passes
over a value in a cell, all the cells with the same value for that attribute are
highlighted, showing the frequency and distribution of this value. Hovering is
even more useful with multiple views to highlight parts linked by some relation.
MatrixExplorer [30] uses two linked visual representations for networks, one
being a node-link diagram and the other an adjacency matrix. When the mouse
hover over an entity in one visual representation, the same entity is highlighted
in the other.

Brushing and Linking. This technique involves the user watching multiple views
related to the same dataset. When the pointer is moved over an item in one
view, all the related items are highlighted in all the views [4, 9].
For multivariate networks, these views can use the same visual representation
or a different one; they can show the same information (e.g., the network topology
as a node-link diagram as well as an adjacency matrix [30]), complementary
information (e.g., the network topology as a node-link diagram and nominal
attributes as lists [28]), or mixed aspects (e.g., the network topology as a node-
link diagram and attributes using parallel coordinates [3,55]). These can be used
to more easily contrast and compare information or variables in distant places
within the network or to see parts of the dataset from different perspectives. The
latter is often used when visual encoding does not allow for viewing all the data
in one visual representation. This is generally caused by data size (too many
data points to show) or data complexity (too many data variables).
Further interaction techniques are often used to augment and enhance the
use of multiple views. Some of these will be described in the visual encodings
section.
 7

Magic Lenses. Magic lenses [7] are “filters, that modify the presentation of
application objects to reveal hidden information, to enhance data of interest,
or to suppress distracting information.” They have been used extensively in
visualization of networks and multivariate data.
Excentric Labeling [20] are like tooltips: they interactively display labels over
dense visualizations such as scatterplots or node-link diagrams. When enabled,
they show a focal region (rectangular or circular) that follow the mouse; all the
items inside the region are labeled outside of the region with a line connecting
each item to its label. Bertini et al. [5] has extended upon this to give better
control of the focal region and visualization of aggregated information on the
focal region.
Jusufi et al. [33] describe lenses for multivariate network that display nodes
as small multidimensional visualizations when they are within the focal area.
They use several visualizations: parallel coordinates, bar chars, and star plots.

Navigation

Panning and zooming Panning and zooming involve changing the visible view-
port over the otherwise unchanged visualized data. These actions are usually
accomplished via standard interactions with common controls like scroll bars
and sliders, hardware like mouse scroll-wheels and track-pads or using multi-
touch at touch tables or tables.
Several navigation techniques have been designed to improve panning and
zooming over large data sets, since this is very cumbersome, often requiring users
to drag the mouse for long distances over the screen. The simplest technique to
overcome that problem is to use overview plus details representations, such as
a bird’s eye view of the visualization in a small window and detailed view in a
large one. The viewport of the detailed view is usually displayed as a rectangle
on the small window, and can usually be manipulated for fast panning.
In graph visualization, topology aware graph navigation allows automatic pan-
ning and zooming in a graph. These actions can be performed directly on the
network structure, such as link sliding [42] or bring-and-go [58]. These techniques
allow the user to quickly find out-of-viewport nodes that are attached to a par-
ticular node, relocate these to be temporarily positioned in their current view
and then allow further navigation from them. The bring-and-go technique can
also be considered as a magic-lens for navigation.

View distortion (single/multiple) View distortion allocates more space to items

of the users’ interest. In particular, fisheye views generally allow people to see
more information at a point of interest. For example, this can reveal detailed
information that was initially smaller than one pixel in size.
For graphs, there are specific distortion techniques, such as Balloon Focus in
a treemap [59] and a guaranteed visibility technique in dendrograms [43] that
allocate more space to the nodes in focus for their detailed inspection. These
techniques allow for multiple foci at the same time.
8

Distortion can be applied also to edges, improving the visibility of items on

the screen. For example, Edge Lenses [66] interactively displace edges under the
pointer in order to avoid overplotting of edges over nodes or edges over each
other. Tominski et al. [58] have proposed two types of lenses to facilitate the
exploration of networks: Local Edge Lens only show edges with vertices inside
the focal region to locally reduce clutter; Bring Neighbors Lens transiently moves
vertices that are connected to vertices in the focal area but not visible in the
viewport at the boundaries of the focal area. Their lenses can also be combined.
Note that the latter technique can be seen as an example of magic lenses.
The view distortion is not always geometric: Semantic zooming changes the
visual representation and level or details according to the zoom level. The in-
teraction technique remains the same as pan & zoom (e.g., using the mouse
wheel or a zoom slider) but the visual effect of zooming is changed. Semantic
zooming [46] involves changing the visual parameters by altering the amount of
detail shown at various levels of zoom. The simpler kind of semantic zooming
consists of showing more details when zooming in, and less when zooming out,
connecting the zoom level to the data aggregation level [19]. This could involve
showing more of a network at the greater zoom depth such as changing graph
aggregation level [17].

3.2 Visual-Structure Level Interactions

Selection Selection interactions alter the visual parameters of the visualization.
They generally result in the most basic form of encoding change in order to
highlight or emphasize areas of the network. Often they modify visual attributes
of the graph entities (e.g., color, size, line width, etc). Selection differs from
the view-level highlighting in that implies a state change at least at the visual-
structure level, sometimes even at the data level. Also, highlighting is transient
and changes implicitly as the pointer moves or the search query changes, whereas
selection is explicitly set on or off.
There are various ways of selecting. For graphs one can select/brush nodes di-
rectly by clicking on them, select nodes according to their network properties [6]
or select items according to network attribute values. The latter is specifically
suitable when analyzing multivariate networks. Moreover, the network struc-
ture can be used for an enhanced highlighting, i.e., not only the selected nodes
are highlighted but also their neighbors or parent/child nodes. This can be ex-
tended with node or edge properties, where only those adjacent/connected nodes
are highlighted which have certain node attribute values. An example is high-
lighting of controlled companies in a shareholding network [57].
McGuffin et al. [40] have described techniques to select subgraphs interac-
tively. In addition to traditional rectangle and lasso selection of nodes, they in-
troduce a special kind of radial menu to further control and extend the selection
of nodes (e.g., extending it by increasing a radius from the current selection: add
nodes at distance 2, 3, etc.) They also introduce a special kind of menu box that
appear transiently to operate on the current selection for visual-structure-level
or data-level operations (e.g., align, change color, change shape, etc.).
 9

Changing mapping of attributes Interactions that change the visual encoding

can also be used to explore and understand various dimensions of the data. An
example of this is changing the visual mapping of attributes, i.e., which attributes
are assigned to which visual attributes such as size and color. Such interaction
should be typically provided in interactive graph visualization systems.
Even considering just classic node-link representations for networks, visual
encodings and styles of these may still vary greatly. Different emphasis can be
given to visual objects, such as by drawing edges faintly using a high level of
transparency or displaying nodes as points without size. These choice can in
turn lead to vastly different visual results for the same data. Hence, interactively
varying such attributes of the visual encoding can be useful to discover different
properties of the data. See [37] some of the more extreme examples.

Network layout Layout-based interactions alter the position of nodes and edges
based on properties of the network. The intent is for the layout to reveal addi-
tional information about the structure of the network.
Examples of layout-altering interactions include positioning nodes and edges
to emphasize similarity, such as using Multidimensional Scaling [36], or by ap-
plying existing automated graph layout algorithms. Interactions to apply layout
changes are typically triggered by changing a layout setting, however layout can
sometimes be adjusted by interacting directly with the network, i.e., dragging
nodes or edges.
Network layout can be calculated solely in dependence of network struc-
ture [26], only in dependence on node properties [6] or a combination of both
network structure and network attributes [34, 52]. The type of layout depends
on the user task. If the user wishes to analyze the relationships between nodes
in the network, a topology-only layout is sufficient. However, if she wishes to an-
alyze the interconnection of network structure and network attributes (e.g., are
people with similar characteristics friends?), a layout that takes both network
structure and network attributes into account is preferable.
Moreover, constraint-based network layout approaches can allow interactive
control and fine-tuning of the layout [16], and may be used in conjunction with
multiple views and semantic zooming to allow interactive browsing and explo-
ration of large multivariate networks [15].
Multiple differing network layouts can be coupled with multiple views and
augmented with brushing and other highlighting techniques to understand the
relations between them [11]. This allows the user to compare and analyze the
network from different perspectives, and detect information which might have
been hidden while using a single layout.

Representation Graphs and multivariate data can be represented visually in var-

ious ways (e.g., node-link diagrams vs. adjacency matrix for graphs; scatterplot
matrices vs. parallel coordinate plots for multivariate data, etc). As one repre-
sentation may not reveal the intended information on the network, the user may
wish to change the representation in order to gain a better view of the data.
This is done using interactions altering visual encoding of parts of the network
10

or present alternative representations such as matrix views, tables, or even a

mixed representation such as in NodeTrix [31].

3.3 Data Level Interactions

Data-based interactions involve selecting which data to show (showing more, less
or completely different data) or manipulating data values (deleting, inserting
data).

Selecting Data for Visualization

Filtering: For large graphs, the whole graph may not be shown on the screen.
The user then can decide either to reduce the size of the displayed data set
(filtering) or to expand on demand the currently shown part of the data set
(adding undisplayed data). Then, data level filtering interaction enables display
of just interesting subsets of the data.
Such interaction can be performed directly in the network visualization (by
selecting nodes to hide) or using a query interface. The query interface can range
from a simple slider for attribute values, a histogram-based filter up to a filtering
via brushing in additional views on the data (multiple views).

Dynamic Querying: Sometimes there can be one or more important variables to

focus on within the visualization. A prominent example is time. The user may
wish to browse through time in the visualization of dynamic graphs. Then it is
useful to provide controls to directly allow the user to quickly move through the
range of possible values. This is analogous to using sliders and other common
controls to provide panning and zooming for the space dimension.

Adding undisplayed data: An alternative way of exploring large graphs is to

show a small part of the graph at the beginning of the analysis process (e.g.,
as a result of a search for interesting nodes) and then expand this selection
on demand [27]. The expansion allows the user to add undisplayed data to the
network. This might be via navigating through the network topology, such as
showing neighboring nodes or connections between nodes on demand [28]. In
hierarchic graphs (trees), one can navigate along the hierarchy and show nodes
on a lower level of hierarchy, or show nodes only a certain level [19].
The number of possible expansions of a graph might be very large. Then
the user may not know which parts of the graph to expand. In such situations,
it is useful to show information on which elements to display when there are
more candidates than there is room to show. Such decisions are often based on
a degree of interest function. Such functions can be calculated in many different
ways (e.g., [23, 27, 29, 39]).
 11

Search: Search-based interactions at the data level are most useful when not all
of the multivariate network data can be shown at once. They allow particular
entities of interest to be extracted and displayed or highlighted from the entire
data set. Specific examples are to:

– search for nodes/edges with certain attribute values;

– search for nodes/edges with certain topological properties;
– search for subgraphs with specific structural properties (motifs) [62]; and
– search for graphs—interactive user interfaces for defining query graphs and
searching for them [61]

Search actions may be performed in various ways. They may involve construc-
tion of textual or graphic queries, may be performed by example, or achieved by
finding similar items to those in a selection drawn or otherwise specified by the
user. Search interactions may result in other data-level changes such as filtering
and adding undisplayed data.

Pivoting: In the case that different variables are represented by different edge
and node types in a heterogeneous network, pivoting is an interaction approach
where the user can visualize a couple of variables at once and switch between
looking at various slices of the data. Usually this involves keeping some common
part of the network visible and as stable as possible during pivot actions, such
as in PivotGraph [64] or PivotPaths [13].

Changing Data In some cases, the user may wish to change the input data
such as data attributes or graph structure. This can be done by direct data
editing in the user interface or by aggregation.

Editing: Multiple different ways to edit graphs exist:

– Graph structure: The user may wish to edit the graph structure: delete or
add nodes or edges. This is usually done directly in the visual interface
by selecting nodes/edges to delete or by drawing new nodes or edges. The
system can then either show these changes directly or can also show impact
of these changes on network structure [62].
– Attributes: In multivariate networks, the user may additionally change at-
tribute values for certain nodes. Moreover, the user can run a specific algo-
rithm, which creates new attribute values that can be explored or used for
navigation in the graph. This includes creating new attributes by combining
existing attributes (such as sum of two attributes) or by creating attributes
describing node or edge topologic attributes (e.g., betweenness centrality).

Aggregation: Large graphs are often simplified by aggregation. Aggregation

merges several nodes and/or edges to so-called supernodes or superedges, where
one supernode represents several nodes and one superedge represents several
edges. The user may choose to see one of the predefined graph levels (pre-defined
12

aggregation) or define the aggregation interactively. Such aggregation can merge

user-selected nodes into one node [2] or can automatically merge nodes based on
user-defined node attributes [64] or on topologic network properties [62]. The ag-
gregation based on selected network attributes is specially useful for multivariate
networks. This allows for variable views on the graph and its structure.

Annotation: Annotation is an interaction where the user can add additional in-
formation to objects in the visualization in order to augment their understanding
of the data and indicate or signpost points of interest. This is analogous to us-
ing notes in order to make sense of complexity, although this is arguably more
valuable when it is done in-place by annotating the network directly. In this way
the annotations cause changes to the data which subsequently allows the user
to search, filter or otherwise interact with the annotations directly.

History and Provenance: Interactive exploration and analysis of large graphs

includes many steps—interaction actions—and feedback loops. The performed
interactions are then difficult to remember and reproduce. This is facilitated
by tracking of user actions. GraphDice [6] records view changes and selection
changes and show them as a set of miniatures. Hovering over a miniature tran-
siently changes the selection to use the one recorded in the history. Clicking on
the miniature sets the view and selection to the recorded one. RelaNet System
tracks and automatically aggregates all user actions [60]. It then shows them to
the user using a graphical representation: a tree whose nodes are visualization
states and edges are actions. The user that click on a node in the tree, which
resumes one of the previous visualization states. The user can then either reply
the actions or start a new exploration path (creating a new branch in the tree).
All actions can be stored, shared and reviewed.
The recorded actions can then be analyzed algorithmically or shown to the
user for their visual inspection. The CZSaw system [35] keeps track of all user
interactions and allows to explore and share them.
The tracking, reproducibility and analysis of user actions is still a large chal-
lenge in visual analytics area. This problem belongs to the more general issue of
analytical provenance7 addressed by systems such as VisTrails [54].

4 Exemplars
To better illustrate effective interaction techniques and methodologies for multi-
variate graphs, we here present four exemplars of existing InfoVis systems that
include such techniques. These exemplars include the GraphDice system by Bez-
erianos et al. [6], the GraphTrail system by Dunne et al. [14], Parallel Node-Link
Bands by Ghani et al. [22], and the state transition networks by Pretorius and
van Wijk [48]. We used the following criteria when selecting these exemplars for
inclusion in the chapter:
7
See http://www.vacommunity.org/AnalyticProvenanceWorkshop for the first work-
shop on this issue
 13

Fig. 2. The GraphDice [6] multivariate visualization tool visualizing an IEEE InfoVis
co-authorship network consisting of both intrinsic and derived attributes. The analyst
is in the process of transitioning between two different node attributes; the transition
is shown as a smooth animation.

– Representative: Our objective was to select exemplars that capture a wide

range of representative interaction techniques;
– Significant: The included examples all provide interaction techniques that
are among the first of their kind;
– Best practices: All exemplars demonstrate best practices in interaction for
multivariate graphs; and
– Familiar: Our selection is by necessity limited by our knowledge, experience,
and preconceptions of the general field of multivariate graph interactions.

In no way do we claim that this set of exemplars is exhaustive or optimal.

There may exist several other InfoVis tools that we could have selected instead
of these four. We only claim that our selection is representative and illustrative.
We use the term “analyst” to refer to a domain specialist performing analysis
tasks with the system, rather than a “data analyst”.

4.1 GraphDice

GraphDice [6] is a multivariate graph visualization tool that supports navigat-

ing in data space similar to the scatterplot matrix navigation proposed in the
ScatterDice [18] tool. The key contribution of GraphDice is the integration of
attribute-based layout with interactive data space navigation, where both in-
trinsic (such as the age, gender, and annual income) as well as derived (layout
position, degrees, and centrality) attributes of actors in a social network form the
data space. This supports a smooth and fluid visual exploration process where
users can seamlessly sculpt their queries across all attributes.
In terms of specific interaction techniques for multivariate graphs, GraphDice
supports the following:
14

Fig. 3. GraphTrail [14] overview of a multivariate co-authorship dataset for the ACM
CHI conference. The screenshot shows examples of the tag cloud, hybrid bar chart,
and matrix chart supported by the tool.

– Smoothly changing visual mapping: The key feature of data space navi-
gation [18] is to smoothly change the mapping of attribute dimensions to po-
sitional (X and Y) visual variables using an animated transition. GraphDice
does not discriminate between intrinsic and computed attributes, thereby al-
lowing the analyst to transition from a geographic or computed graph layout
to other attributes such as degree, centrality, age, gender, income, etc.
– Pivoting: Data space navigation in GraphDice also allows for pivoting a
multivariate graph to study different slices, or facets, of the data. This inter-
action is inspired by PivotGraphs [64], and also incorporates node and link
aggregation to minimize overplotting and to summarize a large number of
data points. Similarly, GraphDice also summarizes multiple time points into
intervals that are visible during pivoting.
– Query sculpting: Query sculpting is a faceted filtering technique that is
closely integrated with the data space navigation and pivoting functionality
in GraphDice. The analyst can use lasso, bounding box, or interval selec-
tion on the main node-link display to create queries in the dataset. These
queries are maintained in a query control box, which also summarizes the
size, distribution, and name of each query. Analysts can then use data space
navigation to pivot the query, allowing them to sculpt it by adding additional
constraints on other attribute dimensions.
 15

4.2 GraphTrail
The GraphTrail [14] visual analytics tool by Dunne et al. supports exploration
of graph data where the nodes and links are both multivariate—containing mul-
tiple attributes, as already prominently discussed in this chapter—as well as
multimodal (called heterogeneous in the paper)—where the nodes or links are
of different types, or modes. The work presents two case studies: (1) publication
data for the ACM CHI conference (Figure 3), where nodes contain attributes
such as year, title, name, locations, and date, and the modes are authors, papers,
and proceedings; and (2) a large-scale archeological graph of artifacts consisting
of 24 different node modes and 35 link modes. The GraphTrail tool supports the
following specific interaction techniques for multivariate graphs:

– Aggregation: The tool presents aggregated views of graphs in self-contained

charts such as bar charts, tag clouds, and tables instead of the raw graph
data as a traditional node-link diagram. The purpose is to use familiar and
readable visual summaries as opposed to the full graph dataset.
– Visual history: While not strictly a multivariate graph interaction tech-
nique, GraphTrail provides an innate visual interaction history by maintain-
ing each exploration branch as a chain, or trail, of connected charts. This
allows the analyst to refer back to the exploration path, which may poten-
tially be branching, at any time.
– Exploratory interactions: The tool supports three specific interaction
techniques for multivariate graph exploration:
– Filtering and merging: Selecting subsets of a dataset for drill-down and
merging disparate subsets into a single chart using direct manipulation;
– Pivoting: Transitioning between different edge and node types (i.e., modes)
to explore multimodal relationship in the graph; and
– Cloning: Duplicating subsets and charts with dependencies to avoid hav-
ing to propagate upstream changes to connected child charts.

4.3 State Transition Networks

Pretorius and van Wijk [48] present a multivariate graph visualization tech-
nique for visual inspection of state transition graphs (Figure 5). Such graphs
are common for complex systems and are often used for design, debugging, and
evaluation. The visual representation is based on separating the different modes
in these state transition graphs and showing two modes at a time, essentially
as a bipartite graph layout with nodes on each side of the display and the links
(and edge labels) connecting the modes in-between. The implementation allows
the user to navigate between which two modes to drill down into. In addition,
the system supports several dedicated multivariate graph interaction techniques:

– Selection and highlighting: A simple but key interaction technique is the

ability to select a node (or a cluster of nodes) in the visual representation,
causing all contained or connected nodes and edges to be highlighted in red.
16

Fig. 4. The parallel node-link bands (PNLBs) [22] technique visualizing a multimodal
NSF funding graph. The parallel coordinate inset is a specific interaction technique
called “open sesame” for drilling down into one or several scalar attributes of a set of
nodes; in this case, it is visualizing the year and amount for funded projects.

This interaction is coupled with appropriate visual representations that also

highlight multivariate attributes in the connecting edges.
– Filtering: In conjunction with the selection technique, the multivariate
graph implementation also supports adding to or subtracting from the se-
lection to further refine the analyst’s exploration.
– Clustering: To cope with the large scale of the state transition graphs, the
prototype implementation supports clustering and aggregating nodes and
edges based on attributes and labels.

4.4 Parallel Node-Link Bands

Similar to GraphTrail, the parallel node-link bands (PNLBs) [22] method is

a graph visualization technique for multimodal and multivariate graphs, i.e.,
graphs where nodes and links not only have multiple attributes, but also belong
to two or more different modes, or types. However, instead of focusing on aggre-
gated charts summarizing the network, PNLBs draw on the work by Pretorius
and van Wijk [48] (discussed above) to retain the node-link visual metaphor
but separates the nodes by their respective mode into specific bands organized
by slicing the viewport vertically (Figure 4). Unlike Pretorius and van Wijk,
however, PNLBs generalizes to any number of bands. However, the technique
only shows links between adjacent bands, suppressing all other link modes to
minimize visual clutter. For this reason, the technique also borrows many ideas
from semantic substrates [52], where node modes are organized into spatially
disjoint substrates. However, PNLBs were designed specifically for multivariate
graph exploration, and supports the following interaction techniques in support
of this goal:
 17

Fig. 5. A multivariate graph visualization designed by Pretorius and Van Wijk [48] of
a state transition graph created by a system analyst.

– View distortion: PNLB bands can be zoomed and panned independently

of each other; furthermore, they can also be designed to support semantic
zooming. One particular use of this is to enable view distortion where a
fisheye function around a selection or the user’s mouse cursor can smoothly
expand and compress the visual marks representing the nodes in the graph.

– Multivariate drill-down: The tool allows the analyst to drill down into
entire bands or individual nodes to uncover the multivariate attributes “hid-
den” in the data. For example, tag cloud and details-on-demand popups
can show summaries or the full details of a node. Furthermore, a specific
interaction technique dubbed “open sesame” (visible in Figure 4) integrate
a parallel coordinate inset within an expanded node band to show quanti-
tative data for those nodes; parallel coordinates were chosen because they
closely mimic the overall visual design of PNLBs, but other chart types can
be integrated as well.

– Multimodal drill-down: Another drill-down option focuses on the topo-

logical nature of the graph by exposing the within-network relations within
the dataset, i.e., the links that connected nodes of the same mode. This is
a necessary mechanism since the PNLBs technique is designed to primarily
show between-mode links for adjacent mode bands.
18

5 Recommendations and Guidelines

When designing or evaluating interactive visualizations for multivariate net-

work data it is useful to consider potential interaction techniques with regard
to their usability. Indeed, this is even more important when testing novel un-
proven interaction techniques. There exists a large body of experience in the
Human-Computer Interaction community regarding the usability of interaction
techniques. In this section we describe some well accepted usability principles
and interaction design guidelines and discuss them within the context of multi-
variate network visualization. The information in this section should provide a
useful lens through which to assess the appropriateness of particular interaction
techniques.
We group these guidelines into three broad categories—Learnability, Flexi-
bility and Robustness—as suggested by Dix et al. [12]. We also draw from other
sources, including Cognitive Dimensions of Notations (CDs) [8, 25], which offers
some useful vocabulary for discussing design and choice of interactions as well
as evaluating the impact a design will have on users.

5.1 Learnability

Learnability describes a set of principles that can be used to determine the

ease with which a new user can begin productive work with the system [12].
This is especially important when designing interactions for multivariate network
visualization since these will often present a large amount of data, and complex
interfaces for exploring the dimensions of the data. Hence, anything that can be
done to help users quickly learn the system and accompanying interactions is
important.
One important aspect of Learnability is Predictability which simply states
that a given interaction should always behave predictably, i.e., exhibit deter-
ministic behaviour. Also, it should exhibit Consistency in that an interaction
that can be performed on one element of the visualization should be able to be
applied to other objects and produce similar results. This means sticking with
established conventions for network and multivariate visualization, such as those
described earlier in this chapter.
Other aspects of Learnability are Familiarity and Generalizability which
deal with creating interfaces and interactions that map as closely as possible to
the real world or similar interfaces the user will already be familiar with. Ideally,
this is done to maximize utilization of users’ past experience. This includes
making use of familiar metaphors where applicable, as discussed at length in
Chapter TODO: Metaphors.
Affordance [44] describes the ability of physical or digital objects to suggest
how they my be interacted with through their appearance. For example, the
handles on a drawer afford the user the ability to pull out the drawer. Similarly,
the appearance of standard GUI controls like sliders and buttons suggest how
you can interact with them.
 19

When designing novel interfaces for interacting with multivariate network

visualizations it might not always be possible to give interactive controls or ele-
ments these obvious affordances due to the complexity of the interface, but the
appearance of controls can still sometimes be enhanced in subtle ways to hint at
their intended use, or at least at the possibly of interaction. For example, with
hyperlinks in web documents being colored, it is not always obvious what effect
clicking a link with have, but the user knows it will do something and can make
an educated guess based on the link text and surrounding context. Affordance is
related to the discoverability of interactive capabilities in the interface; with the
profusion of graphical entities shown by visualization, it is always tempting to
provide interactions contextual to specific entities, but without affordance, the
chance that a novice user will discover it by chance is very low. When design-
ing user interface components or assigning contextual interactions to graphical
entities, ask yourself if they adequately express their role to the user?
Avoiding Hidden Dependencies (CDs) [8,25] means that the link between
connected items should be visible and obvious to the user. This can be a problem
in any system with filtering and search where the visualization may just show
a subset of the results. Ideally, if the information linking this information is
significant the visualization can show the smallest subset of the network that
links the nodes and edges in the search results. This becomes more of a problem
when an interaction has a surprising result due to such a hidden dependency.
Supporting Progressive Evaluation (CDs) [8, 25] means allowing users to
be able to take a break at any time and take stock of their progress so far. This
is especially important for exploratory tasks involving novice users.

5.2 Flexibility

Cognitive load [41] is based on the fact that humans can hold relatively little
information in short term memory—famously, seven plus/minus two pieces of
information. As a result, we must consider the amount of information that needs
to be retained in working memory in order to effectively work with a system
or interface. As much as possible there is a need to alleviate the user having
to commit unnecessary information to memory. In terms of visualizations of
multivariate networks, this means relevant data should be highly visible and
the interface should as much as possible make it clear at a glance the data and
attributes that the users is looking at, as well as how they reached this point,
and how they can return to earlier points of exploration. Memory required for
any complex processing task should be able to be delegated to system where it
should be presented in a fashion that is easy for the users to understand.
Fitts’s law [21] is a model that describes the act of pointing. It says that the
time taken to rapidly move a pointing device to a target object is proportional
to the size of the target and the distance to it. The implication of this for
interaction and interface design is that the most frequently used controls should
be the closest and largest. In the case of visualizations for multivariate networks,
a prerequisite for answering the question of where to put the controls would
20

be understanding the kinds of tasks that users were going to perform most
commonly with the visualization, since context matters [1].
Visualizations should have low Viscosity (CDs) [8, 25], i.e., common tasks
should be able to be accomplished with a minimal number of actions or effort
on the part of the users.
Abstraction (CDs) [8, 25] involves providing shortcuts to the user in order
to facilitate them working efficiently with logical sets of the data at once. This
is often vital in multivariate network data, since the aim is to allow the user to
manipulate the visualization at the level of a particular attribute or dimensions
of the data rather than forcing them to interrogate the properties associated
with individual nodes and edges themselves. Abstractions should be used where
possible, since them simply many tasks and help with understanding the network
and associated variables.
Terseness and Diffuseness (CDs) [8, 25] state that it is important to ded-
icate appropriate amounts of display space to the various elements of the visu-
alization, rather than devoting too much or too little space to them. This may
seem obvious, but question the space that is being used to show various elements
of the data set.
An important general quality to strive for in designing interactive visualiza-
tions of multivariate networks is providing good Guidance, both in terms of
dimensions and graph structure. That is, when the user can not currently see
some particular dimension of the data or a section of network, we would like to
let them know this information exists and also give them some estimation of the
importance of the non-visible information.

5.3 Robustness
Direct Manipulation [50] describes interaction that is performed directly on
objects and provides continuous, fast feedback in response to change. Another
way of thinking about this is providing continuity and thus avoiding abrupt
changes that could potentially confuse the user. When an action is not being per-
formed by the user directly, it can sometime be smoothly animated to achieve
a similar effect. Direct manipulation behavior maps more closely to how objects
behave in the real world and has a few important benefits. It allows users to
quickly determine or predict the final outcome of their action, it allows them
to more easily realize when performing an action would lead to unintended con-
sequences, and to more easily reverse an action. It also allows users to reach a
desired end result in a single action that would otherwise require several actions
when seeing the effects on the system only after each of the individual actions
had been performed.
Recoverability [12] is an important robustness principle for most user in-
teraction. It suggests that our visualization system should easily allow the user
to undo any action they have made in error. As we mentioned before responsive
interactions such as direct manipulation approaches also help in this regard.
Premature Commitment (CDs) [8, 25] means users should not have to
make any decision before they have adequate information to base it upon. This
 21

can often be solved by providing a flexibility to the system where the user can
reach a particular result of view of the data via multiple paths, rather than just
a single specific sequence of actions. Also, the user should have the ability to try
things out without committing to them (Provisionality).
Error-Proneness (CDs) [8,25] describes the ability of the system to induce
errors from the user and not protect them from these mistakes. When evaluating
an interface to a complicated multivariate network visualization we must consider
what is being show to the user. Could they easily mistake or confuse some aspect
of the visualization and reach an incorrect inference or conclusion? We want to
avoid this.

Finally, it should be noted that creating effective interactive visualizations for

multivariate network data is a difficult task that combines the inherent complex-
ity of navigating large graphs with understanding and exploring multiple dimen-
sions. As noted by Pike et al. [47], it is important to remember that the purpose
of interaction is to enable an analytic discourse during which users build, test and
refine knowledge. Hence, the design of new interaction approaches should care-
fully focus on the likely aims, intentions and actions of the user. They should also
be formally evaluated through user studies and have their effectiveness tested
with real users.

6 Challenges and Vision

This chapter has discussed and classified various interactive techniques for mul-
tivariate and network visualization. It has explored their use in several effective
multivariate network visualization systems, and has described guidelines for de-
signing successful interaction approaches. Here we conclude the chapter by out-
lining what we see as the major unsolved challenges in this space, and offering
our thoughts on where research in this field might head in coming years.
Comprehend landscape: The first big challenge we see is that visualiza-
tion research and systems often give good individual detailed views of particular
facets of the data, but don’t necessarily offer a visual interface that allows un-
derstanding of the entire information landscape at a high level. The difficulty is
in giving enough of a sense of what the data is, conveying its meaning, as well
as hinting at the dimensions or facets of the data that could be worth exploring
in greater depth.
Similarly, in the case where multivariate network visualizations make use of
multiple views there is a challenge in providing elegant and simple mechanisms
to manage these views. This is important since a single view will never be ade-
quate to explore large multivariate networks and users will spend significant time
controlling, comparing and navigating between views. Additionally, particular
application domains often require their own unique multiple view configurations
that are still a challenge to build interactively [65].
Provenance: We think a significant unaddressed challenge is supporting
provenance in multivariate network visualizations. This involves assisting the
22

user in tracking the exploration process, and the history of their interactions
with the system. This is important since these actions form a critical part of
the analysis process. While the information visualization community is aware of
the importance of recording and showing which interactions were performed and
when [24, 45, 47, 60], these methods either do not support multivariate networks
or only to a small extent (e.g., selecting variables for visual mapping and search).
Ad hoc design: Currently, many multivariate network visualizations are
built by extending upon existing systems using traditional interfaces. This is
generally not a good idea when the interfaces are required to encapsulate so
much complexity. Ideally, we would like to see more approaches utilising prin-
cipled design from the beginning. That is, designing interfaces and interactions
specifically to support required application rather than just bolting additional
controls and complexity onto existing interfaces. Specifically, this means tech-
niques should be designed based-on principles and guidelines like those presented
in Section 5, but also be given formal user testing to evaluate their effectiveness.
This is not to say that we can’t have a set of general purpose visualization
components and reuse these, but in order for such objects to be useful they will
require standardized, consistent interaction. We need not just a common vocab-
ulary and behavior, but also an understanding of their specificity and efficiency
for particular uses.
Evaluation: There is a real need for formal evaluation of interaction ap-
proaches used for visualization of multivariate networks. There are a range of
issues and concerns here. Firstly, there needs to be more consideration of the
tasks for which the visualizations are to be used, since this can often be ne-
glected. For such an analysis of tasks, see Chapters TODO: Tasks. Do we
know what users really need when dealing with multivariate data? How does
this change when they are exploring vs. checking a hypothesis vs. using the
visualization to convince another person of some fact. When is interaction sup-
portive and can it become a burden? We know there needs to more evaluation
of techniques supporting exploration and offering guidance so that the users do
not get “lost” in the space.
While this sort of evaluation and experience partly exists for networks, the
approaches and studies do not explicitly include and deal with multivariate net-
works. We need to enhance our understanding around faceted exploration along
both multivariate and network data at the same time, since much of what we
know applies only to exploration within a single dimension. Finally, there needs
to be more work on automatic and semi-automatic identification of important
nodes and dimensions for directing the user during analysis tasks, as well as
subsequent evaluation of such approaches.
Temporal data: Another challenge is in dealing with multivariate networks
with a temporal dimension. Obviously time can be treated as just another vari-
able. Although given that humans have a particular understanding of time, it
can be useful to leverage this familiarity and treat it specially when designing
interactive visualizations. This is explored in detail in Chapter TODO: Tem-
poral.
 23

Scalability: Dealing with scalability is obviously a major challenge. While

interaction can help with some of the issues, it is often complicated by scale. For
example, can we fully show both dimensions and network data, or must we reduce
the complexity of the data and its presentation? Is interaction in real time still
possible on large data sets? Scalability considerations for multivariate network
visualization will be discussed in detail in Chapter TODO: Scalability.
Emerging hardware: Multivariate network visualization could potentially
make use of emerging hardware such as multitouch tables and tablets, contact-
less input devices (Microsoft Kinect, Leap Motion Controller), wall-sized dis-
plays, and even 3D stereoscopic displays or immersive cave environments. Of
course, these technologies are exciting to use or witness for the first time but
there are a lot of unknowns surrounding their use. Do more pixels help solve the
problems we have discussed? Can we benefit from extra dimension in 3D without
the typical downsides, such as users becoming disoriented or lost? Can we utilize
navigational multitouch gestures? Can multimodal input provide extra benefits
specifically for navigating multidimensional data? How can we support collabo-
rative data exploration of multivariate networks between multiple simultaneous
local or remote users? Then there is the possibility of building novel gadgets or
devices specifically for interacting with this sort of data and visualization.
The new hardware settings provide opportunities but are challenging too.
Should we build one visualization application for a specific piece of hardware
(e.g., a multitouch tablet) or can we design interactions that will adapt to a
range of input devices and output capabilities? The description of the input side
of interactive applications is still at its infancy in HCI, and the management
of powerful interaction approaches is relatively new in the visualization field.
Adapting visualization interactions to the new setups is therefore a huge chal-
lenge [32]. Research in this area will likely start by exploring specific hardware
configurations and progressively evolve towards some unification for classes of
technologies over time.

Acknowledgements
The genesis and planning of this chapter took place at the Schloss Dagstuhl semi-
nar “Information Visualization—Towards Multivariate Network Visualization” held in
May 2013. We wish to thank Guy Melançon for his useful contributions during these
discussions.

References
1. Appert, C., Beaudouin-Lafon, M., Mackay, W.E.: Context matters: Evaluating
interaction techniques with the CIS model. In: Fincher, S., Markopoulos, P., Moore,
D., Ruddle, R. (eds.) People and Computers XVIII Design for Life, pp. 279–295.
Springer London (2005)
2. Archambault, D., Munzner, T., Auber, D.: GrouseFlocks: Steerable exploration of
graph hierarchy space. IEEE Transactions on Visualization and Computer Graph-
ics 14(4), 900–913 (2008)
24

3. Auber, D., Archambault, D., Bourqui, R., Lambert, A., Mathiaut, M., Mary, P.,
Delest, M., Dubois, J., Melançon, G.: The Tulip 3 Framework: A Scalable Soft-
ware Library for Information Visualization Applications Based on Relational Data.
Technical Report RR-7860, INRIA (Jan 2012)
4. Becker, R.A., Cleveland, W.S.: Brushing scatterplots. Technometrics 29(2), 127–
142 (May 1987)
5. Bertini, E., Rigamonti, M., Lalanne, D.: Extended excentric labeling. In: Proceed-
ings of the 11th Eurographics / IEEE - VGTC conference on Visualization. pp. 927–
934. EuroVis’09, Eurographics Association, Aire-la-Ville, Switzerland, Switzerland
(2009)
6. Bezerianos, A., Chevalier, F., Dragicevic, P., Elmqvist, N., Fekete, J.D.:
GraphDice: A system for exploring multivariate social networks. Computer Graph-
ics Forum 29(3), 863–872 (2010)
7. Bier, E.A., Stone, M.C., Pier, K., Buxton, W., DeRose, T.D.: Toolglass and magic
lenses: the see-through interface. In: Proceedings of the 20th annual conference on
Computer graphics and interactive techniques. pp. 73–80. SIGGRAPH ’93, ACM,
New York, NY, USA (1993)
8. Blackwell, A.F., Britton, C., Cox, A.L., Green, T.R.G., Gurr, C.A., Kadoda, G.F.,
Kutar, M., Loomes, M., Nehaniv, C.L., Petre, M., Roast, C., Roe, C., Wong, A.,
Young, R.M.: Cognitive dimensions of notations: Design tools for cognitive tech-
nology. In: CT ’01: Proceedings of the 4th International Conference on Cognitive
Technology. pp. 325–341. Springer (2001)
9. Buja, A., McDonald, J.A., Michalak, J., Stuetzle, W.: Interactive data visualization
using focusing and linking. In: Proceedings of the 2nd Conference on Visualization.
pp. 156–163. VIS ’91, IEEE Computer Society Press, Los Alamitos, CA, USA
(1991)
10. Card, S., Mackinlay, J., Shneiderman, B.: Readings in Information Visualization:
Using Vision to Think. Morgan Kaufmann Publishers (1999)
11. Collins, C., Carpendale, S.: VisLink: Revealing relationships amongst visualiza-
tions. IEEE Transactions on Visualization and Computer Graphics 13(6), 1192–
1199 (2007)
12. Dix, A., Finlay, J., Abowd, G.D., Beale, R.: Human-Computer Interaction (3rd
Edition). Prentice-Hall, Inc. (2003)
13. Dörk, M., Riche, N.H., Ramos, G., Dumais, S.: PivotPaths: Strolling through
faceted information spaces. IEEE Transactions on Visualization and Computer
Graphics 18(12), 2709–2718 (2012)
14. Dunne, C., Riche, N.H., Lee, B., Metoyer, R., Robertson, G.: GraphTrail: An-
alyzing large multivariate, heterogeneous networks while supporting exploration
history. In: Proceedings of the ACM Conference on Human Factors in Computer
Systems. pp. 1663–1672 (2012)
15. Dwyer, T., Marriott, K., Schreiber, F., Stuckey, P., Woodward, M., Wybrow, M.:
Exploration of networks using overview+detail with constraint-based cooperative
layout. IEEE Transactions on Visualization and Computer Graphics 14(6), 1293–
1300 (2008)
16. Dwyer, T., Marriott, K., Wybrow, M.: Dunnart: A constraint-based network di-
agram authoring tool. In: Proceedings of the International Conference on Graph
Drawing. pp. 420–431. Springer (2009)
17. Elmqvist, N., Do, T.N., Goodell, H., Henry, N., Fekete, J.D.: ZAME: Interactive
large-scale graph visualization. In: Proceedings of the IEEE Pacific Symposium on
Visualization. pp. 215–222 (2008)
 25

18. Elmqvist, N., Dragicevic, P., Fekete, J.D.: Rolling the dice: Multidimensional visual
exploration using scatterplot matrix navigation. IEEE Transactions on Visualiza-
tion and Computer Graphics 14(6), 1141–1148 (2008)
19. Elmqvist, N., Fekete, J.D.: Hierarchical aggregation for information visualization:
Overview, techniques, and design guidelines. IEEE Transactions on Visualization
and Computer Graphics 16(3), 439–454 (2010)
20. Fekete, J.D., Plaisant, C.: Excentric labeling: dynamic neighborhood labeling for
data visualization. In: Proceedings of the SIGCHI conference on Human Factors
in Computing Systems. pp. 512–519. CHI ’99, ACM, New York, NY, USA (1999)
21. Fitts, P.M., Peterson, J.R.: Information capacity of discrete motor responses. Jour-
nal of Experimental Psychology 67, 103–112 (1964)
22. Ghani, S., chul Kwon, B., Lee, S., Yi, J.S., Elmqvist, N.: Visual analytics for
multimodal social network analysis: A design study with social scientists. IEEE
Transactions on Visualization and Computer Graphics (2013), to appear
23. Gladisch, S., Schumann, H., Tominski, C.: Navigation recommendations for ex-
ploring hierarchical graphs. In: 9th International Symposium on Visual Computing
(ISVC’13). p. to appear. Advances in Visual Computing, Springer (2013)
24. Gotz, D., Zhou, M.X.: Characterizing users’ visual analytic activity for insight
provenance. Information Visualization 8(1), 42–55 (2009)
25. Green, T.R.G.: Cognitive dimensions of notations. In: Sutcliffe, A., Macaulay, L.
(eds.) Proceedings of the 5th Conference of the British Computer Society, Human-
Computer Interaction Specialist Group - People and Computers V. pp. 443–460.
Cambridge University Press, New York, NY, USA (1989)
26. Hachul, S., Jünger, M.: Large-graph layout algorithms at work: An experimental
study. Journal of Graph Algorithms and Applications 11(2), 345–369 (2007)
27. van Ham, F., Perer, A.: “Search, Show Context, Expand on Demand”: Supporting
large graph exploration with degree-of-interest. IEEE Transactions on Visualiza-
tion and Computer Graphics 15(6), 953–960 (2009)
28. Heer, J., Boyd, D.: Vizster: Visualizing online social networks. In: IEEE Symposium
on Information Visualization. pp. 32–39 (2005)
29. Heer, J., Card, S.K.: DOITrees revisited: scalable, space-constrained visualization
of hierarchical data. In: Proceedings of the ACM Conference on Advanced Visual
Interfaces. pp. 421–424 (2004)
30. Henry, N., Fekete, J.D.: MatrixExplorer: a dual-representation system to explore
social networks. IEEE Transactions on Visualization and Computer Graphics
12(5), 677–684 (2006)
31. Henry, N., Fekete, J.D., McGuffin, M.J.: NodeTrix: a hybrid visualization of so-
cial networks. IEEE Transactions on Visualization and Computer Graphics 13(6),
1302–1309 (2007)
32. Isenberg, P., Carpendale, S., Bezerianos, A., Henry, N., Fekete, J.D.: CoCoNut-
Trix: collaborative retrofitting for information visualization. IEEE Comput. Graph.
Appl. 29(5), 44–57 (Sep 2009)
33. Jusufi, I., Dingjie, Y., Kerren, A.: The network lens: Interactive exploration of mul-
tivariate networks using visual filtering. Information Visualisation, International
Conference on 0, 35–42 (2010)
34. Jusufi, I., Kerren, A., Zimmer, B.: Multivariate network exploration with jaun-
tynets. In: Proceedings of: 17th International Conference on Information Visuali-
sation (IV ’13). No. 17 (2013), to appear
35. Kadivar, N., Chen, V., Dunsmuir, D., Lee, E., Qian, C., Dill, J., Shaw, C., Wood-
bury, R.: Capturing and supporting the analysis process. In: Visual Analytics Sci-
ence and Technology, 2009. VAST 2009. IEEE Symposium on. pp. 131–138 (2009)
26

36. Kruskal, J., Wish, M.: Multidimensional Scaling. No. no. 11 in 07, SAGE Publica-
tions (1978)
37. Lima, M.: Visual Complexity: Mapping Patterns of Information. Princeton Archi-
tectural Press (2011)
38. Liu, Z., Navathe, S.B., Stasko, J.T.: Network-based visual analysis of tabular data.
In: IEEE VAST. pp. 41–50 (2011)
39. May, T., Steiger, M., Davey, J., Kohlhammer, J.: Using signposts for navigation in
large graphs. Computer Graphics Forum 31(3), 985–994 (2012)
40. McGuffin, M.J., Jurisica, I.: Interaction techniques for selecting and manipulat-
ing subgraphs in network visualizations. IEEE Transactions on Visualization and
Computer Graphics 15(6), 937–944 (Nov 2009)
41. Miller, G.A.: The magical number seven, plus or minus two: some limits on our
capacity for processing information. Psychological Review 63, 81–97 (1956)
42. Moscovich, T., Chevalier, F., Henry, N., Pietriga, E., Fekete, J.D.: Topology-aware
navigation in large networks. In: Proceedings of the ACM Conference on Human
Factors in Computing Systems. pp. 2319–2328 (2009)
43. Munzner, T., Guimbretière, F., Tasiran, S., Zhang, L., Zhou, Y.: Treejuxtaposer:
scalable tree comparison using focus+context with guaranteed visibility. In: ACM
SIGGRAPH 2003 Papers. pp. 453–462. SIGGRAPH ’03, ACM, New York, NY,
USA (2003)
44. Norman, D.A.: The design of everyday things. Basic Books (2002)
45. North, C., Chang, R., Endert, A., Dou, W., May, R., Pike, B., Fink, G.: Ana-
lytic provenance: process+interaction+insight. In: Extended Abstracts of the ACM
Conference on Human Factors in Computing Systems. pp. 33–36 (2011)
46. Perlin, K., Fox, D.: Pad: an alternative approach to the computer interface. In:
Proceedings of the 20th annual conference on Computer graphics and interactive
techniques. pp. 57–64. SIGGRAPH ’93, ACM, New York, NY, USA (1993)
47. Pike, W.A., Stasko, J., Chang, R., O’Connell, T.A.: The science of interaction.
Information Visualization 8(4), 263–274 (2009)
48. Pretorius, A.J., van Wijk, J.J.: Visual inspection of multivariate graphs. In: Pro-
ceedings of the Eurographics / IEEE - VGTC conference on Visualization. pp.
967–974 (2008)
49. Robinson, A.H.: The thematic maps of Charles Joseph Minard. Imago Mundi 21,
95–108 (1967)
50. Shneiderman, B.: Direct manipulation: A step beyond programming languages.
IEEE Computer 16(8), 57–69 (1983)
51. Shneiderman, B.: The eyes have it: a task by data type taxonomy for information
visualizations. In: Proc. Int. Symp. on Visual Languages. pp. 336–343 (1996)
52. Shneiderman, B., Aris, A.: Network visualization by semantic substrates. IEEE
Transactions on Visualization and Computer Graphics 12(5), 733–740 (2006)
53. Shneiderman, B., Plaisant, C.: Designing the User Interface — Strategies for Ef-
fective Human-Computer Interaction (5. ed.). Addison-Wesley (2010)
54. Silva, C., Freire, J., Callahan, S.: Provenance for visualizations: Reproducibility
and beyond. Computing in Science Engineering 9(5), 82–89 (2007)
55. Smoot, M.E., Ono, K., Ruscheinski, J., Wang, P.L., Ideker, T.: Cytoscape 2.8:
new features for data integration and network visualization. Bioinformatics 27(3),
431–432 (2011)
56. Spenke, M., Beilken, C., Berlage, T.: Focus: the interactive table for product com-
parison and selection. In: Proceedings of the 9th annual ACM symposium on User
interface software and technology. pp. 41–50. UIST ’96, ACM, New York, NY, USA
(1996)
 27

57. Tekušová, T., Kohlhammer, J.: Visual analysis and exploration of complex corpo-
rate shareholder networks. In: Proceedings of the SPIE Conference on Visualiza-
tion and Data Analysis. pp. 68090F–68090F. International Society for Optics and
Photonics (2008)
58. Tominski, C., Abello, J., van Ham, F., Schumann, H.: Fisheye tree views and
lenses for graph visualization. In: Proceedings of the Internationl Conference on
Information Visualization. pp. 17–24 (2006)
59. Tu, Y., Shen, H.W.: Balloon focus: a seamless multi-focus+ context method for
treemaps. IEEE Transactions on Visualization and Computer Graphics 14(6),
1157–1164 (2008)
60. von Landesberger, T., Fiebig, S., Bremm, S., Kuijper, A., Fellner, D.: Handbook
of Human Centric Visualization, chap. Interaction Taxonomy for Tracking of User
Actions in Visual Analytics Applications, p. to appear. Springer (2013)
61. von Landesberger, T., Bremm, S., Bernard, J., Schreck, T.: Smart query definition
for content-based search in large sets of graphs. In: Proceedings of EuroVAST. pp.
7–12. European Association for Computer Graphics (Eurographics), Eurographics
Association, Goslar (2010)
62. von Landesberger, T., Görner, M., Rehner, R., Schreck, T.: A system for interactive
visual analysis of large graphs using motifs in graph editing and aggregation. In:
Proceedings of the Vision Modeling Visualization Workshop. pp. 331–339 (2009)
63. von Landesberger, T., Kuijper, A., Schreck, T., Kohlhammer, J., van Wijk, J.,
Fekete, J.D., Fellner, D.W.: Visual Analysis of Large Graphs: State-of-the-Art and
Future Research Challenges. Computer Graphics Forum 30(6), 1719–1749 (2011)
64. Wattenberg, M.: Visual exploration of multivariate graphs. In: Proceedings of the
ACM Conference on Human Factors in Computing Systems. pp. 811–819 (2006)
65. Weaver, C.: Building highly-coordinated visualizations in improvise. Information
Visualization, IEEE Symposium on 0, 159–166 (2004)
66. Wong, N., Carpendale, S., Greenberg, S.: Edgelens: An interactive method for
managing edge congestion in graphs. In: Information Visualization, 2003. INFOVIS
2003. IEEE Symposium on. pp. 51–58. IEEE (2003)
67. Yi, J.S., Kang, Y.a., Stasko, J.T., Jacko, J.A.: Toward a deeper understanding of
the role of interaction in information visualization. IEEE Transactions on Visual-
ization and Computer Graphics 13(6), 1224–1231 (2006)