Visual Exploration of Relationships between Document Clusters

Ilir Jusuﬁ

1,2

, Andreas Kerren

, Jiayi Liu

and Bj

orn Zimmer

Linnaeus University, Department of Computer Science, ISOVIS Group, V

axj

o, Sweden

Institute for Data Analysis and Visualization, Department of Computer Science, University of California, Davis, U.S.A.

Keywords:

Network Visualization, Multivariate Data, Clustering, Document Visualization, Text Visualization, Interac-

tion, Visual Analytics.

Abstract:

The visualization of networks with additional attributes attached to the network elements is one of the ongoing

challenges in the information visualization domain. Such so-called multivariate networks regularly appear

in various application ﬁelds, for instance, in data sets which describe friendship networks or co-authorship

networks. Here, we focus on networks that are based on text documents, i.e., the network nodes represent

documents and the edges show relationships between them. Those relationships can be derived from common

topics or common co-authors. Attached attributes may be speciﬁc keywords (topics), keyword frequencies,

etc. The analysis of such multivariate networks is challenging, because a deeper understanding of the data

provided depends on effective visualization and interaction techniques that are able to bring all types of in-

formation together. In addition, automatic analysis methods should be used to support the analysis process

of potentially large amounts of data. In this paper, we present a visualization approach that tackles those

analysis problems. Our implementation provides a combination of new techniques that shows intra-cluster

and inter-cluster relations while giving insight into the content of the cluster attributes. Hence, it facilitates

the interactive exploration of the networks under consideration by showing the relationships between node

clusters in context of network topology and multivariate attributes.

1 INTRODUCTION

Researchers across different domains of science deal

with an increasing number of network data. A com-

plicating fact is that real life networks usually have at-

tributes associated with their elements. The data types

of those attributes can be either homogeneous or het-

erogeneous. Attributes of a co-authorship network,

for example, with nodes representing documents and

edges indicating if documents share a common topic

(keywords) could be considered homogeneous since

the keywords are always quantiﬁed as connections be-

tween documents. In other words, all attributes (key-

words) have the same data type depending on what

the edges actually represent: either Boolean (con-

tained in a document or not) or Integer (frequency

in a document). A more complex social network, in-

volving age, gender, salary, favorite sports team, and

other preferences, however, would be considered het-

erogeneous. Homogeneous attributes have the advan-

tage that they can all be visualized using the same

metaphor (visual representation). The visualization of

such multivariate networks is one of the ongoing chal-

lenges in the information visualization domain [Ker-

ren et al., 2013].

Often, researchers who analyze such networks do

not have any inherent understanding about the struc-

ture or the values of the data at hand. This problem

has been increased by the growing amount of data

produced with the help of new technologies. Some-

times it is even hard to ask interesting research ques-

tions or forming a simple hypothesis on speciﬁc as-

pects of the data. Therefore, we often need a start-

ing point for our visual exploration. One solution

to ﬁnd such an initial state may be Shneiderman’s

mantra “overview ﬁrst, zoom and ﬁlter, then details-

on-demand” [Shneiderman, 1996]. Another possibil-

ity is given by automatic analyses, such as methods

for unsupervised clustering (k-means, etc.) to make

sense of large input data sets [Mirkin, 2005].

Let us assume that we applied a clustering algo-

rithm on our multivariate network to group the nodes

on the basis of their attributes. It is often insufﬁ-

cient to show only relationships between clusters as

there may exist signiﬁcant connections between clus-

ters and individual data elements. In consequence, the

visualization of the underlying network together with

the attributes and the derived cluster data may help

researchers revealing interesting data patterns. Here,

the visual analysis of the links between the network

195

Jusuﬁ I., Kerren A., Liu J. and Zimmer B..

Visual Exploration of Relationships between Document Clusters.

DOI: 10.5220/0004754301950203

In Proceedings of the 5th International Conference on Information Visualization Theory and Applications (IVAPP-2014), pages 195-203

ISBN: 978-989-758-005-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

nodes within a speciﬁc cluster may be an important

help to discover possible correlations between the at-

tributes and the network topology. Additionally, the

presence of strong relationships between nodes from

different clusters may indicate the existence of unique

connection types that are unrelated to the concrete at-

tribute values (as such relationships do not exist in

one single cluster only). A straightforward approach

to visualize attribute-based clusters in networks is to

aggregate clustered nodes into super-nodes. However,

this approach both neglects to include intra-cluster

(i.e., cluster-local) relationships and to identify indi-

vidual cluster members that are highly connected to

other clusters.

In this paper, we present a visualization prototype

that is used to visually analyze scientiﬁc conference

articles which form a co-authorship network. Ad-

ditional metadata has been used to cluster the docu-

ments, and the results are represented through our vi-

sualization and interaction techniques. Our visualiza-

tion and interaction techniques supported by the pro-

totype tackle the problem of representing intra-cluster

relationships in networks while providing additional

insight into the attributes of the elements within the

clusters. Thus, users are able to visualize intra-cluster

relationships, inter-cluster relationships, and relation-

ships between clusters and individual elements with

the help of different interaction techniques. Although

similar approaches that deal with such tasks exist (for

instance the work of [Henry et al., 2007]), our pro-

totype introduces a combination of techniques that

shows intra-cluster and inter-cluster relations while

giving insight into the content of the cluster attributes.

The rest of the paper is structured as follows: a

brief discussion of related work is presented in Sec-

tion 2. The description of our main visualization and

interaction ideas are discussed in Section 3. Technical

details of our prototype are highlighted in Section 4

followed by two use cases in Section 5. Finally, we

conclude the paper with a brief summary and a de-

scription of our planned efforts for the future.

2 RELATED WORK

Various systems visualize multivariate network data

and the topology of the network itself. [Borisjuk et al.,

2005] employ small diagrams as nodes instead of sim-

ple circles or rectangles in metabolic networks. Each

diagram represents experimental data that is related

to the node. To handle the clutter that could be in-

troduced if the embedded diagrams cause too many

overlaps, magic lenses can be used to show different

multivariate attributes on demand [Jusuﬁ et al., 2010].

GraphTrail [Dunne et al., 2012] is another system to

analyze multivariate data and supports the exploration

of heterogeneous networks. It also introduces an in-

teraction history to allow users to refer back to previ-

ous exploration steps. These approaches do not pro-

vide any clustering mechanisms for data analysis.

Another approach to avoid clutter is to use so-

called Semantic Substrates [Shneiderman and Aris,

2006]. Nodes are placed in non-overlapping regions

based on their attributes. A related technique de-

veloped by [Wattenberg, 2006] uses a grid-layout to

show the relationships between node attributes and

links. [Pretorius and van Wijk, 2008] arrange edge la-

bels in a list and place rectangular regions contain-

ing source and target nodes at each side. The re-

gions are partitioned according to the attributes of the

nodes and are connected to corresponding edge labels

via straight lines. Based on this idea, parallel node-

link bands [Ghani et al., 2013] separate nodes into

multiple regions and visualize them similar to a par-

allel coordinates plot. Previous work of [Pretorius

and van Wijk, 2007] introduces hierarchical cluster-

ing based on node attributes to place nodes in spe-

ciﬁc regions. Related techniques are presented by

[Archambault et al., 2009, Archambault et al., 2008].

However, the underlying graph topology is not com-

pletely visible or hard to perceive with the aforemen-

tioned techniques. Most of these tools place the nodes

in distinct regions which could be considered as a

clustering or grouping. However, none of them pro-

vides insight into the attributes of the clusters/groups

or into their context. A slightly different technique

is introduced by [Jusuﬁ et al., 2013]. They use an

attribute-driven layout to steer node positioning based

on multivariate attributes, while still showing the gen-

eral network structure. Similar works have been pub-

lished before [Bezerianos et al., 2010, Wu and Takat-

suka, 2006]. One drawback of those tools is that the

visualization of clusters is either not possible or lim-

ited.

There are different techniques to visualize text-

based attributes. Tag clouds [Kaser and Lemire,

2007] are one approach to visualize the frequency of

words in a document or the keywords of a document

collection, and there are several very popular tools

which extend the original work, such as Wordle [Vie-

gas et al., 2009], ManiWorlde [Koh et al., 2010], or

ProjCloud [Paulovich et al., 2012]. Our prototype

uses a simple version of tag clouds to show common

keywords of documents that were clustered together.

WordBridge [Elmqvist and Ebert, 2011] combines tag

clouds with node-link diagrams in order to show the

details of relationships between the entities in text col-

lections. They use tag clouds to represent nodes and

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

196

Figure 1: A screenshot of our implementation. The toolbox is placed on the left hand side: the ﬁrst button is used for opening

data ﬁles, the next two buttons provide switching between two alternative cluster layouts, and the radio buttons (numbered

from 3 to 7) specify the number of desired clusters. The histogram shows the edge weight distribution, i.e., the x-axis

represents the weight and the y-axis the corresponding number of edges. It can be used to select ranges of edge weights. In

the main drawing area on the right hand side, the visualization of ﬁve distinct clusters is shown together with external edges

routed through the center (circular cluster layout). The edges are highlighted in blue color, if they are hovered (note that the

edge bundling algorithm may superimpose edges).

edges, but do not employ clustering techniques on the

data. A brief overview of state-of-the-art text visu-

alization techniques can be found in [Kerren et al.,

2012]. Another recent tool for text document analy-

sis is iVisClustering [Lee et al., 2012]. The system

focuses on the interactive analysis of clustered docu-

ments and uses multiple coordinated views to assist

the user reﬁning the clustering results. While iVis-

Clustering is used for text document collections only,

our system could also be applied to general multivari-

ate networks.

Even with all the presented techniques at hand, the

visualization of multivariate networks is still a chal-

lenge. Our prototype supports the interactive explo-

ration of clustered multivariate data and the relation-

ships inside of and between clusters. Instead of show-

ing a complete node-link drawing of the multivari-

ate network at the beginning of the analysis, we de-

cided to show the resulting cluster graph directly. In

order to do this more efﬁciently, we have developed

a novel cluster visualization approach that combines

tag clouds with cluster graphs (Figure 1). As such,

our new layout integrates the original nodes and edges

into the visualization. The design and features of this

new approach are explained in the following section.

3 VISUALIZATION AND

INTERACTION TECHNIQUES

In this section, we present our main interaction and

visualization techniques. We start with a description

of the data set used to test our approach.

3.1 Example Data Set

Our tool was tested with a Jigsaw data set contain-

ing metadata for every IEEE InfoVis conference paper

from 1995 to 2011 as well as VAST papers from 2006

to 2011 [Jigsaw, 2011]. The papers themselves are

represented by (network) nodes, and co-authorship is

VisualExplorationofRelationshipsbetweenDocumentClusters

197

represented by links, i.e, if two papers share an author,

then their node representations are connected with an

edge. Both together form a co-authorship network.

The weight of an edge corresponds to the number

of shared authors. So-called concept terms [Jigsaw,

2011] describe the paper’s contents and are used as

node attributes. They are metadata that were identi-

ﬁed by parsing the paper titles and abstracts. There-

fore, each (node) attribute in this data set has only

boolean values: one and zero, i.e., either a term exists

in a speciﬁc paper or not. In addition, the Jigsaw data

set provides more attributes, such as the publication

year for each paper.

3.2 Document and Cluster Visualization

After loading the data, the user is prompted to select

the desired attributes to be clustered and visualized

through a dialog box. In this way, one can ﬁlter out

uninteresting attributes in advance. Additionally, the

user can specify one speciﬁc node attribute (e.g., the

publication year) from the list of already chosen at-

tributes whose values should be mapped to the node

representations directly.

As stated in [Kerren and Schreiber, 2012], there

are several ways to integrate automatic analysis re-

sults into visualizations. For simplicity and the other

reasons described in Section 1, we decided to start

with the automatic analysis and to use its results for

the subsequent visualization (other architectural de-

signs of the tool are possible and useful depending on

the analysis tasks). After k-means clustering (the user

is able to choose an appropriate k), the network nodes

are arranged on the screen according to their cluster

membership. By referring to our example data set,

each paper is represented by a network node drawn

in form of a donut slice shape and placed on a cir-

cular layout. In consequence, all nodes that belong

to the same cluster build a node ring as depicted in

Figure 2. The nodes in this ﬁgure have been ordered

based on the preselected attribute value as described

above. Therefore, the saturation of the node color

(green) in the screenshot changes gradient-wise. Plac-

ing clusters in a circular layout is a common practice

in graph drawing, with one of the earliest works pre-

sented by [Kar et al., 1988]. This approach supports

the easy identiﬁcation of clusters while showing the

overall network structure.

The weighted edges which represent co-

authorship between documents are drawn by using

curved lines with a yellow-to-red color gradient. The

gradient visualizes the number of authors shared

between two documents (edge weights) with yellow

being the lowest number of common co-authors (at

least one) and red the highest number (normalized

Figure 2: A cluster representation showing the nodes ar-

ranged radially. They are colored in green, and the color

saturation represents the value of any preselected attribute

(in this concrete case the publication year). The concept

terms of all nodes (papers) are arranged in the center. The

edges between these nodes are routed around the terms in

order to avoid overlaps.

Figure 3: Clusters visualized by using our free cluster lay-

out method (k = 3). External edges are partitioned into sev-

eral bundles in order to connect neighbored clusters (node

rings).

over a maximum number of common co-authors).

The edges have been routed inside of the node ring to

facilitate the drawing of the tag cloud that represents

the concept terms. This tag cloud gives insight into

the main concepts described in this particular group

of documents. Since we are dealing with text data in

this case, tag clouds are an intuitive choice of visual

metaphor. However, the cluster attributes can be

visualized using any appropriate metaphor without

modifying the general framework.

Figure 1 shows the graphical user interface (GUI)

of our implemented prototype. On the left hand side, a

toolbox featuring different controls is located. On the

right hand side, the main visualization view is visible.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

198

Figure 4: A list of all documents is shown on the left. The selected document is highlighted in orange, while the documents

connected to it are highlighted in yellow. Those tags that are not associated with the selected document are faded out in the

corresponding node ring (the one at the top in our screenshot; k = 3).

It is possible to specify the number of desired clus-

ters through a user control. The clustered nodes are

drawn as previously described and placed in a circu-

lar layout with the edges bundled through the center

of the view. Because the edges originate and end at

data elements on the perimeter of the cluster nodes,

we retain information about both cluster connections

and element connections. Such inter-cluster edges are

called external edges while the edges within a single

cluster are called internal edges in the remainder of

this paper. External edges are color-coded based on

the edge weight in the same way as internal edges.

The edge routing within the circular arrangement

of the cluster rings might introduce visual clutter in

case the number of nodes and edges is very high.

Therefore, an interactive free cluster layout method

was added to avoid such issues. Here, the cluster rings

can be arranged manually, see Figure 3. We imple-

mented a different edge routing method for this layout

to reduce the visual complexity. All outgoing (exter-

nal) edges from a particular cluster are radially routed

outside the node ring to a speciﬁc control point that

leads them to other clusters and vice versa. In this

way, the two control points of neighbored node rings

(clusters) establish a kind of a beachhead. All pass-

ing external edges are bundled together in due con-

sideration of a speciﬁc order: the edges with the high-

est weight (i.e., links with most shared authors) are

placed in the middle of the bundle. Users can switch

between our two layouts by clicking the correspond-

ing icons located in the upper-right corner of the tool-

box (Figure 1). More details about the edge routing

for both layouts are given in Section 4.

3.3 Interaction Techniques

Several standard interaction techniques are supported

to facilitate the visual exploration in our prototype.

If the user selects a node, then the internal edges of

the corresponding cluster and all the external edges

that are not connected to the selected node are ﬁltered

out. This is helpful for the identiﬁcation of related

documents, i.e., all papers that share at least one au-

thor in our case. A tooltip with the document name

appears once the user hovers the mouse over a corre-

sponding node. Our software can also display a list

with all documents as shown in Figure 4. The se-

lected document is highlighted in orange. Tags not

found in the selected document will be faded out; the

present tags give a quick overview of the content of

the selected document. In the current version of our

implementation, the user has to follow the edges to

identify related papers and to select these to get the

corresponding concept terms. Our approach can be

easily extended to support the selection of several re-

lated nodes which would give the user the possibility

to see, for instance, common concept terms visualized

by the tag clouds.

Edge selection can be cumbersome when dealing

with a lot of edges routed through narrow areas as

VisualExplorationofRelationshipsbetweenDocumentClusters

199

seen in Figure 1. Therefore, we implemented a spe-

ciﬁc selection possibility when several edges are close

to the mouse cursor when a click action is performed.

Initially, all edges close to the mouse cursor are high-

lighted in blue as shown in Figure 1; optionally a

tooltip shows the number of highlighted edges. After

clicking, only these highlighted edges remain visible

and all other edges are ﬁltered out. This operation can

be repeated consecutively until only one edge can be

selected. This usually does not require many clicks

until the view becomes uncluttered. The tooltip of the

remaining edge ﬁnally displays the shared authors be-

tween the two linked documents. A left mouse click

is used to cancel the selection (all edges are shown

again). Additionally, the user can ﬁlter out edges with

the help of the interactive histogram in the toolbox,

which shows the edge weight distribution (Figure 5)

and also mimics range sliders [Williamson and Shnei-

derman, 1992]. All edges with weights outside the

speciﬁed range are ﬁltered out, reducing clutter. Our

interactive histogram provides insight about possible

interesting range selections that could be made simi-

lar to data visualization sliders [Eick, 1994]. As ex-

pected, most of the documents share exactly one au-

thor in our data set (cf. Figure 5).

Figure 5: A histogram shows the edge weight distribution.

It can also be used as selection tool of weight ranges. In this

particular example, the majority of edges have the weight 1,

i.e., the documents share one author.

Our visualization approach was tested using a net-

work comprised of 578 nodes. It is obvious that scal-

ability issues are introduced with the increasing num-

ber of nodes and/or clusters. We have not performed

experiments yet to ﬁnd out what the maximum num-

ber of nodes and/or clusters is that users can analyze

with our approach. This will be part of our future

work.

4 TECHNICAL ASPECTS

In this section, a description of used tools and details

about our edge routing approaches are provided. Our

prototype is implemented in Java to guarantee a cer-

tain level of platform independence. For clustering

the attribute data, we used the k-means algorithm al-

though other clustering algorithms could be used as

well. We have integrated the clustering functionality

with the help of the trickl-cluster library which offers

several clustering algorithms that could be used alter-

natively [Trickl, 2011].

Quadratic B

ezier curves [Farin, 2002] were used

for the implementation of the edge routing for the cir-

cular cluster layout. A circular area through which

all external edges should pass is deﬁned in the center

of the layout as exempliﬁed in Figure 6 (the dashed

circle). The ﬁrst couples of anchor points (P

, P

)

and (P

, P

) close to the node rings make sure that

the edges are connected to the nodes at a perpendic-

ular angle. Each node ring has a different diameter.

Therefore, the edges are drawn in such a way that they

do not stretch over a speciﬁc radius from the current

ring while going through the middle in order to avoid

edge crossings with neighboring rings. This is imple-

mented by using the anchor points P

and P

. Each

of them is speciﬁed as point of intersection between

a node ring tangent originating from the center of the

dashed circle and the circle itself.

Figure 6: The diagram exempliﬁes the edge routing ap-

proach in our circular cluster layout.

Figure 7: The diagram describes the edge routing in the free

cluster layout.

As indicated in the previous section, straight seg-

ments are used to connect the clusters in the free clus-

ter layout. However, the edges are routed around the

node rings in order to reach the straight line. The ap-

proach is schematically displayed in Figure 7. First,

a straight line is drawn from the node, i.e., from point

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

200

to point P

. From there, a circle segment is drawn

around the entire node ring up to the point P

. This

point and P

serve as beachheads between the two

rings and are connected with a new straight line. From

this point, the algorithm uses the same steps as ex-

plained for the ﬁrst node ring. More implementation

details can be found in the thesis [Liu, 2013].

5 USE CASE

Our tool facilitates the local exploration of the con-

nectivity between documents grouped into clusters

and helps to get insight into the node attributes—in

our case concept terms of the documents belonging

to a particular cluster. For instance, a high number

of internal edges with high weight values might indi-

cate speciﬁc patterns, e.g., that groups of people have

been published together a high number of papers on

a speciﬁc topic as exempliﬁed in the following. Dur-

ing the interactive exploration of our data set, we no-

ticed that only one group of three authors published

two papers in the same year. This can be seen af-

ter ﬁltering out all papers in the same cluster with

less than three authors, compare Figure 8. One of

the internal edges is very short which gives us a hint

that these two papers were published in the same year

(because the two corresponding nodes have the same

green color). After the selection of one of the papers,

the unrelated concept terms are faded out. This gives

us insight into the content of the speciﬁed document

in context of the corresponding cluster. Here, the most

prominent concept term is “graph” as it is the largest

word in the tag cloud. The selected document has

this concept term and is not faded out. This shows

both: it has a strong connection to most of the cluster

members, and the authors of the aforementioned pa-

pers published work on graphs twice in the same year.

The selected paper is “MoleView: An Attribute and

Structure-Based Semantic Lens for Large Element-

Based Plots” by Telea, Hurter, and Ersoy published

at the InfoVis conference in 2011. The same au-

thors joined by Paulovich and Cantereiro published

the other paper in 2011 titled “Skeleton-Based Edge

Bundling for Graph Visualization” at the same con-

ference.

Another use case demonstrates the application of

our approach to show external relationships of nodes

in different clusters. By using our tool, we can easily

see that one of the nodes has a lot of external edges;

some of them have a high weight which connects this

node to two separate clusters. Again, edges with a

low number of co-authors were ﬁltered out (Figure 9).

After selecting the corresponding paper (cf. the high-

Figure 8: A cluster with a selected node in orange color. It

shows that a relatively high number of co-authors published

two papers within the same area in the same year—the black

arrow points to the corresponding edge. The publication

year is represented by a green color gradient. Concept terms

that are not contained in the selected paper were faded out.

Figure 9: The screenshot shows the relation of a speciﬁc

selected document with other documents belonging to other

clusters (k = 3).

lighted paper in Cluster A), we are able to see that four

co-authors published papers together that were placed

in different clusters. The tag cloud indicates concept

terms of the selected paper. By clicking on the related

papers in the other node rings (Clusters A and B), it

turns out that the papers in the other rings have almost

no similarities in terms of tag clouds except for the tag

“interaction” (a pretty general concept term) which is

present in all the clusters involved in our case. The

selected paper in Figure 9 (Cluster A) is “Comparing

Different Levels of Interaction Constraints for Deriv-

ing Visual Problem Isomorphs” where the “visual an-

alytics” concept term is prominently displayed. This

is an indicator for the VAST conference where the pa-

per was presented in 2010. The authors of this pa-

per are Ribarsky, Chang, Dou, Ziemkiewicz, Jeong,

VisualExplorationofRelationshipsbetweenDocumentClusters

201

Harrison, Ryan, and Wang. From these, Wang, Dou,

Ribarsky, and Chang co-authored the paper “Parallel-

Topics: A Probabilistic Approach to Exploring Doc-

ument Collections” shown in Cluster B and related

to text and document visualization. The other paper

“Evaluating the Relationship Between User Interac-

tion and Financial Visual Analysis”—shown on the

lower right hand side (Cluster C) of Figure 9—was

written by Jeong, Dou, Chang, Ribarsky, Lipford, and

Stukes and published in 2008 at the VAST confer-

ence as well. This shows that a group of authors has

worked together in various topics of visual analytics

since their papers are found in different clusters, but

they are related to each other.

6 CONCLUSIONS AND FUTURE

WORK

A set of new techniques for the visual analysis of mul-

tivariate network clusters has been presented in this

paper. They facilitate the exploration of clustered data

by (1) showing the cluster content through the use of

tag clouds and (2) giving insight into the underlying

network through the use of two different cluster lay-

out techniques and edge routing algorithms. The anal-

ysis process is enriched with various interaction tech-

niques, such as interactive edge ﬁltering.

There are several improvements that could further

strengthen our prototype. At the moment, only one at-

tribute can be mapped to the nodes. By simply adding

more donut slices on top of the existing node ring and

using different colors, it is possible to visualize more

attributes. This approach might be limited, because

the size of the node rings will increase. Another im-

provement would be to introduce standard zooming

and panning. At the current state, our implementa-

tion does not allow this due to the Java 2D graphics

renderer performance. Therefore, we have to port the

application to OpenGL. The edge routing algorithm

in the circular cluster layout can be improved by rout-

ing the edges directly between cluster neighbors. This

will help to reduce clutter in the center, but it might

introduce a lot of additional edge crossings around

the clusters. Another possibility for clutter reduction

is to display the most interesting edges ﬁrst (subject

to user deﬁned parameters) and to add more on de-

mand. The current interaction possibilities have to be

extended, for instance, by multiple selection of nodes

in order to provide comparisons of the concept terms

between related documents inside and outside of the

same cluster. Last but not least, we have to evaluate

our approach with respect to usability, efﬁciency, and

scalability.

ACKNOWLEDGEMENTS

We would like to thank Alfredo Gimenez for carefully

proof-reading the ﬁnal version of this paper.

REFERENCES

Archambault, D., Munzner, T., and Auber, D. (2008).

GrouseFlocks: Steerable exploration of graph hier-

archy space. Visualization and Computer Graphics,

IEEE Transactions on, 14(4):900 –913.

Archambault, D., Munzner, T., and Auber, D. (2009).

TugGraph: Path-preserving hierarchies for browsing

proximity and paths in graphs. In Visualization Sym-

posium, 2009. PaciﬁcVis ’09. IEEE Paciﬁc, pages 113

–120.

Bezerianos, A., Chevalier, F., Dragicevic, P., Elmqvist, N.,

and Fekete, J.-D. (2010). GraphDice: A system for

exploring multivariate social networks. Computer

Graphics Forum (Proc. EuroVis 2010), 29(3):863–

872.

Borisjuk, L., Hajirezaei, M.-R., Klukas, C., Rolletschek, H.,

and Schreiber, F. (2005). Integrating data from biolog-

ical experiments into metabolic networks with the dbe

information system. In Silico Biol, 5(2):93–102.

Dunne, C., Riche, N. H., Lee, B., Metoyer, R., and Robert-

son, G. (2012). GraphTrail: Analyzing large multi-

variate, heterogeneous networks while supporting ex-

ploration history. In Proceedings of the ACM Confer-

ence on Human Factors in Computer Systems, pages

1663–1672.

Eick, S. G. (1994). Data visualization sliders. In Proceed-

ings of the 7th annual ACM symposium on User inter-

face software and technology, UIST ’94, pages 119–

120, New York, NY, USA. ACM.

Elmqvist, N. and Ebert, D. S. (2011). WordBridge: Using

composite tag clouds in node-link diagrams for visual-

izing content and relations in text corpora. 2011 44th

Hawaii International Conference on System Sciences,

pages 1–8.

Farin, G. (2002). Curves and Surfaces for CAGD: A Prac-

tical Guide. Morgan-Kaufmann, 5th edition.

Ghani, S., Chul Kwon, B., Lee, S., Yi, J., and Elmqvist,

N. (2013). Visual analytics for multimodal social net-

work analysis: A design study with social scientists.

In IEEE Transactions on Visualization and Computer

Graphics.

Henry, N., Fekete, J.-D., and McGufﬁn, M. J. (2007). Node-

Trix: a hybrid visualization of social networks. IEEE

Transactions on Visualization and Computer Graph-

ics, 13:1302–1309.

Jigsaw (2011). Visual Analytics for Exploring and Under-

standing Document Collections. http://www.cc.

gatech.edu/gvu/ii/jigsaw/datafiles.html,

last accessed: 2013-02-15.

Jusuﬁ, I., Dingjie, Y., and Kerren, A. (2010). The network

lens: Interactive exploration of multivariate networks

using visual ﬁltering. In Information Visualisation

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

202

(IV), 2010 14th International Conference, pages 35 –

42.

Jusuﬁ, I., Kerren, A., and Zimmer, B. (2013). Multivariate

network exploration with JauntyNets. In Proceedings

of the 17th International Conference on Information

Visualisation (IV ’13), pages 19–27.

Kar, G., Madden, B., and Gilbert, R. (1988). Heuristic lay-

out algorithms for network management presentation

services. Network, IEEE, 2(6):29–36.

Kaser, O. and Lemire, D. (2007). Tag-cloud drawing: Al-

gorithms for cloud visualization. Proceedings of Tag-

ging and Metadata for Social Information Organiza-

tion (WWW ’07).

Kerren, A., Kyusakova, M., and Paradis, C. (2012). From

culture to text to interactive visualization of wine re-

views. In Marchese, F. T. and Banissi, E., editors,

Knowledge Visualization Currents: From Text to Art

to Culture, chapter 5, pages 85–110. Springer, Oxford.

Kerren, A., Purchase, H., and Ward, M. O. (2013). Informa-

tion Visualization - Towards Multivariate Network Vi-

sualization (Dagstuhl Seminar 13201). Dagstuhl Re-

ports, 3(5):19–42.

Kerren, A. and Schreiber, F. (2012). Toward the role of

interaction in visual analytics. In Proceedings of

the Winter Simulation Conference, WSC ’12, pages

420:1–420:13. Winter Simulation Conference.

Koh, K., Lee, B., Kim, B., and Seo, J. (2010). ManiWor-

dle: Providing ﬂexible control over Wordle. IEEE

Transactions on Visualization and Computer Graph-

ics, 16:1190–1197.

Lee, H., Kihm, J., Choo, J., Stasko, J., and Park, H. (2012).

iVisClustering: An interactive visual document clus-

tering via topic modeling. Computer Graphics Forum,

31:1155–1164.

Liu, J. (2013). Visualization of relationships in clustered

text data. Master’s thesis, Linnaeus University, De-

partment of Computer Science, V

axj

o, Sweden.

Mirkin, B. (2005). Clustering for Data Mining: A Data Re-

covery Approach. Chapman & Hall/CRC, Boca Ra-

ton, FL, USA.

Paulovich, F. V., Toledo, F. M. B., Telles, G. P., Minghim,

R., and Nonato, L. G. (2012). Semantic wordiﬁca-

tion of document collections. Comp. Graph. Forum,

31(3):1145–1153.

Pretorius, A. J. and van Wijk, J. J. (2007). Bridging the

semantic gap: Visualizing transition graphs with user-

deﬁned diagrams. Computer Graphics and Applica-

tions, IEEE, 27(5):58 –66.

Pretorius, A. J. and van Wijk, J. J. (2008). Visual inspec-

tion of multivariate graphs. Comput. Graph. Forum,

27(3):967–974.

Shneiderman, B. (1996). The eyes have it: a task by

data type taxonomy for information visualizations. In

Proceedings of the IEEE Symposium on Visual Lan-

guages, 1996, pages 336–343.

Shneiderman, B. and Aris, A. (2006). Network visualiza-

tion by semantic substrates. IEEE Transactions on Vi-

sualization and Computer Graphics, 12:733–740.

Trickl (2011). trickl-cluster. https://github.com/

trickl/trickl-cluster, last accessed: 2013-02-

15.

Viegas, F., Wattenberg, M., and Feinberg, J. (2009). Partic-

ipatory visualization with Wordle. IEEE Transactions

on Visualization and Computer Graphics, 15:1137–

1144.

Wattenberg, M. (2006). Visual exploration of multivari-

ate graphs. In CHI ’06: Proceedings of the SIGCHI

conference on Human Factors in computing systems,

pages 811–819, New York, NY, USA. ACM.

Williamson, C. and Shneiderman, B. (1992). The dynamic

homeﬁnder: evaluating dynamic queries in a real-

estate information exploration system. In Proceedings

of the 15th annual international ACM SIGIR confer-

ence on Research and development in information re-

trieval, SIGIR ’92, pages 338–346, New York, NY,

USA. ACM.

Wu, Y. and Takatsuka, M. (2006). Visualizing multivariate

network on the surface of a sphere. In Asia Paciﬁc

Symposium on Information Visualisation, pages 77–

83.

VisualExplorationofRelationshipsbetweenDocumentClusters

203