2011 IEEE Social Computing Nodexl: Group-In-A-Box

Group-In-a-Box Layout for Multi-faceted Analysis
of Communities
Eduarda Mendes Rodrigues*, Natasa Milic-Frayling †, Marc Smith‡, Ben Shneiderman§, Derek Hansen¶
*
Dept. of Informatics Engineering, Faculty of Engineering, University of Porto, Portugal
eduardamr@acm.org
†
Microsoft Research, Cambridge, UK
natasamf@microsoft.com
‡
Connected Action Consulting Group, Belmont, California, USA
marc@connectedaction.net
§
Dept. of Computer Science & Human-Computer Interaction Lab
University of Maryland, College Park, Maryland, USA
ben@cs.umd.edu
¶
College of Information Studies & Center for the Advanced Study of Communities and Information
University of Maryland, College Park, Maryland, USA
dlhansen@umd.edu

Abstract—Communities in social networks emerge from One particularly important aspect of social network
interactions among individuals and can be analyzed through a analysis is the detection of communities, i.e., sub-groups of
combination of clustering and graph layout algorithms. These individuals or entities that exhibit tight interconnectivity
approaches result in 2D or 3D visualizations of clustered among the wider population. For example, Twitter users who
graphs, with groups of vertices representing individuals that regularly retweet each other’s messages may form cohesive
form a community. However, in many instances the vertices groups within the Twitter social network. In a network
have attributes that divide individuals into distinct categories visualization they would appear as clusters or sub-graphs,
such as gender, profession, geographic location, and similar. It often colored distinctly or represented by a different vertex
is often important to investigate what categories of individuals
shape in order to convey their group identity.
comprise each community and vice-versa, how the community
structures associate the individuals from the same category.
In addition to the clusters that emerge from the network
Currently, there are no effective methods for analyzing both structure, individuals in a social network are often divided
the community structure and the category-based partitions of into categories that reflect specific attributes. For example,
social graphs. We propose Group-In-a-Box (GIB), a meta- members of the Twitter community may be categorized
layout for clustered graphs that enables multi-faceted analysis based on the number of followers they have, the location
of networks. It uses the treemap space filling technique to from which they tweet, or the date they joined Twitter. Such
display each graph cluster or category group within its own attributes may be useful to gain further insights about the
box, sized according to the number of vertices therein. GIB community. Thus, in addition to detecting communities
optimizes visualization of the network sub-graphs, providing a based on the network structure, it is important to enable
semantic substrate for category-based and cluster-based analysis of the social graph along various attributes of the
partitions of social graphs. We illustrate the application of GIB individuals. At the same time, none of the widely adopted
to multi-faceted analysis of real social networks and discuss layouts for visualizing clustered graphs accommodate such
desirable properties of GIB using synthetic datasets. multi-faceted analysis.
Our research fills that gap by extending the work on
Keywords—network visualization; group-in-a-box; layout; network visualization with semantic substrates [20]. We
meta-layout; force-directed; communities; clustering; semantic introduce Group-In-a-Box (GIB), a meta-layout for clustered
substrates.
graphs that enables multi-faceted analysis of networks. GIB
uses the treemap space filling technique to create graphs for
I. INTRODUCTION individual categories, thus providing a semantic substrate for
Network structures appear in many contexts, from laying out each group. Furthermore, we demonstrate the use
biological systems to communications networks. With the of the method to analyze communities, i.e., graph clusters by
recent proliferation of social media services such as Twitter assigning them specific spatial regions, applying layout
and Facebook, public awareness and usage of social network algorithms to reveal the local cluster structure, and observing
data have increased. This led to a growing need for the categories that their members fall into. In both cases, we
comprehensible visualizations of complex networks that arrive at well separated sub-graphs that clearly reveal the
enable exploratory analysis.

intra-group network structures and the attributes of their where such flows primarily follow the shortest available
members. path.
Our experiments illustrate the application of the GIB The Clauset-Newman-Moore (CNM) algorithm [3], on
multi-faceted analysis to a social network that arises from the the other hand, is a hierarchical agglomeration algorithm for
voting activities of U.S. Senators. Additionally, we use detecting community structure, which is computationally
synthetic networks, generated according to pseudo-random, more efficient than Girvan-Newman and similar algorithms.
small world and preferential attachment network models, to Its running time for a network with n vertices and m edges is
demonstrate that the GIB layout exhibits desirable properties O(md log n), where d is the depth of the hierarchy. However,
for visualizing vertices that connected different clusters. In in practice, the application of CNM is limited to medium-
social networks such individuals are often referred to as size networks, up to half a million vertices. Wakita and
boundary spanners and serve as connectors, forming Tsurumi [23] show that this inefficiency is caused by
communication pathways through which information can be merging communities in an unbalanced manner. They
exchanged across communities. Thus, GIB is equally useful propose simple heuristics to achieve a balanced merge and
for gaining insights about the role of boundary spanners scale up to social networks with millions of users.
within their own community. Based on our experiments, we
recommend generalization of the GIB meta-layout to B. Graph Layouts
accommodate multi-faceted analysis involving hierarchical A common approach to visualizing networks is to use
classification schema and hierarchical clustering algorithms. force-directed layout of vertices and edges based on models
In the following sections, we introduce the graph of physical systems in order to arrive at optimal graph
clustering and layout algorithms used in our study and structures. The Fruchterman-Reingold (FR) layout [9]
describe the GIB layout in detail. Through empirical belongs to that class of algorithms. It treats vertices as steel
observations of specific datasets and network visualization rings and edges as springs between them, observing two
methods, we explore various aspects of the multi-faceted forces: an attractive force that is analogous to a spring force
analysis. We conclude with the discussion of related work abiding by Hooke’s law and a repulsive force that acts as an
and guidelines for devising algorithms for multi-faceted electrical force between charged particles, similar to
analyses, including the need for readability criteria to Coulomb’s law. The algorithm minimizes the energy of the
optimize the utility of the GIB-like layouts. system by moving the vertices and changing the forces
between them until the system reaches an equilibrium state.
II. BACKGROUND AND PROBLEM DEFINITION This layout algorithm is useful for visualizing large
Community detection algorithms aim to identify undirected networks and, generally, creates overall satisfying
cohesive, tightly connected sub-graphs within a network. layouts, placing vertices of the same cluster in the proximity
These structural clusters can be analyzed by considering both of each other. However, some local areas of the graph may
intra-cluster and inter-cluster links. Next, we introduce the still be sub-optimally laid out.
graph clustering and layout algorithms that we used in our Harel and Koren (HK) proposed a fast multi-level graph
experiments and analysis. layout algorithm to achieve better visualizations [11]. Their
A. Detecting Communities through Network Clustering approach involves a two-phase method that recursively
coarsens the graph to arrive at its multi-level representation.
The notion of a community refers to a subset of vertices First, the graph is embedded in a high dimensional space and
among which the distribution of edges is denser than with then projected onto a 2-D plane using principal components
other vertices in the network. There are numerous methods analysis.
for detecting communities, including hierarchical clustering,
graph partitioning by maximizing specific criteria such as C. Faceted Analysis of Communities
network modularity, and similar. In practice, social network analysis (SNA) is multi-
The Girvan–Newman algorithm [7] takes a divisive faceted. Often it is important to analyze the roles of
hierarchical approach. It uses edge betweenness as weights in individuals that arise from their interaction within and across
the divisive process. Community boundaries are detected by communities. However, in order to gain further insights,
progressively removing from the network the edges with the SNA may explore additional, non-structural attributes that
highest betweenness, re-calculating the betweenness of the are not considered by the clustering algorithm. Indeed, the
remaining edges at each step. If a network contains loosely individuals may belong to different categories based on
connected communities, e.g., linked by a few inter-cluster gender, profession, age, location, etc. It is often useful to
edges, then all shortest paths must pass through one of these observe the distribution of these attributes across the
few edges. Thus, such inter-cluster edges will have high community structure. However, to support the analysis that
betweenness and the associated vertices will have high vertex takes into account both the detected community structure and
betweenness centrality [8], a useful measure of the vertex’ the categories that individual belong to, we need to address
influence on the information flow, especially in networks two specific issues:

(a) (b)
Figure 1. (a) Harel-Koren (HK) fast multi-scale layout of a clustered network of Twitter users, using color to differentiate among the vertices in different
clusters. The layout produces a visualization with overlapping cluster positions. (b) Group-in-a-Box (GIB) layout of the same Twitter network: clusters are
distributed in a treemap structure that partitions the drawing canvas based on the size of the clusters and the properties of the rendered layout. Inside each
box, clusters are rendered with the HK layout.

1) There are no layouts that enable flexible visualizations of represent trees through nested rectangular regions. A
categories and links across categories that arise from the rectangular area is subdivided into a set of rectangles that
underlying network structure. represent the top level vertices in the tree hierarchy. This
process continues recursively, creating rectangles that
2) Clustering and layout algorithms are typically applied represent each level in the tree through alternate vertical and
separately. As a result, the graph layout does not take into horizontal subdivision of the rectangles.
account cluster membership of vertices and causes two GIB was implemented as part of the NodeXL network
effects: (i) occlusion of clusters within the graph layout and analysis tool 1 . NodeXL supports computation of network
(ii) loss of information about the structure of individual metrics, such as closeness and betweenness centrality, and
clusters and inter-cluster connectivity. mapping of vertex properties onto visual characteristics, such
The former is illustrated in Figures 1 (a) and Figure 6 (a) as size, color, and shape of the vertices [10][21]. We use
where the CNM clustering algorithm is combined with the NodeXL to generate all the graphs presented in the paper.
HK layout. The clusters overlap, as indicated through ‘color
bleeding’ throughout the network visualization. IV. GIB METHOD FOR MULTI-FACETED ANALYSIS
In this section we demonstrate the use of GIB in multi-
III. GIB LAYOUT ALGORITHM faceted analysis of the U.S. Senate voting patterns in 20072.
We propose to support multi-faceted network analysis by
generalizing the approach of network visualization with A. Dataset
semantic substrates [20]. We present the Group-In-a-Box The U.S. Senate co-voting network was created from the
(GIB) layout that simultaneously captures clustering and data that connects senators to one another based on the
categorization of vertices in a network graph. The GIB number of times they voted the same way (i.e., both in favor
algorithm, or both against a bill). The network graph includes 98
- Uses the treemap space filling technique [1], [15], [19] to vertices and 9506 weighted edges, each representing the
partition the graph canvas into regions of varying sizes, percentage of voting agreement between a pair of senators.
in which individual clusters or category groups are The graph is complete, i.e., all the senators are connected to
displayed. each other. In order to capture strong relationships among the
senators, we focus on edges between senators whose level of
- Allows a choice of layout algorithms for optimizing the agreement is above 50%. The dataset also includes attributes
layout of the sub-graphs within each region. about the senators, such as name, party affiliation, the state
- Enables assignment of visual properties to vertices and they represent, and the number of votes.
edges within and across clusters and category groups.
We choose the treemap approach for the rich information
it conveys about the relative size of individual clusters or 1
NodeXL is freely available at http://nodexl.codeplex.com.
categories (e.g., see Figure 1 (b)) and for its extensibility to 2
Data originally provided by Chris Wilson of Slate magazine available in
hierarchical structures. Indeed, treemaps are designed to the NodeXL template format at http://casci.umd.edu/NodeXL_Teaching.

(a) (b)

Figure 2. The 2007 U.S. Senate co-voting network graph, obtained with the Fruchterman-Reingold (FR) layout. Vertices colors represent the senators’
party affiliations (blue: Democrats; red: Republicans; orange: Independent) and their size is proportional to betweenness centrality. Edges represent
percentage of agreement between senators: (a) above 50%; (b) above 90%.

(a) (b)
Figure 3. The 2007 U.S. Senate co-voting network graph, visualized with the GIB layout. The group in each box represents senators from a given U.S.
region (1: South; 2: Midwest; 3: Northeast; 4: Mountain; 5: Pacific) and individual groups are displayed using the FR layout. Vertices colors represent the
senators’ party affiliations (blue: Democrats; red: Republicans; orange: Independent) and their size is proportional to betweenness centrality. Edges
represent percentage of agreement between senators: (a) above 50%; (b) above 90%..

B. GIB Application to U.S. Senate Co-voting Network by larger vertices), one may want to explore additional facets
The FR layout of the U.S. Senate co-voting network of the data to gain further insights about the voting patterns.
yields the graph shown in Figure 2 (a). In this visualization, We use GIB to perform such analysis. First, we group the
the vertices are colored according to the senators’ party States into five regions: Northeast, Midwest, South,
affiliation and sized according to the vertex betweenness Mountain, and Pacific. We then use these regions to divide
centrality. Further filtering of the edges, to retain those with senators into five groups, one group per region. Using the
weight above 90%, yields the graph shown in Figure 2 (b), same edge filtering thresholds as before, the GIB layout
showing a strong voting agreement among the senators of the results in Figure 3. GIB reveals further information about the
Democratic party. cohesiveness of the two parties across different U.S. regions.
Although this analysis already reveals interesting facts For example, the Republican senators exhibit more voting
about the party cohesion and boundary spanners (represented cohesion in the southern states than the Democrats. Such
observation was not possible in the full graph visualization.

(a) (b) (c)
Figure 4. Small-world network graph visualization obtained with the Harel-Koren layout, after clustering the graph with the Clauset-Newman-Moore
community detection algorithm (5 clusters). (a) Full graph with 500 vertices colored according to the cluster membership. (b) GIB showing the structural
properties of the individiual clusters. (c) GIB layout of the same 5 clusters showing inter-cluster edges.

V. GIB APPLICATION TO COMMUNITY CLUSTERS its own box, increasing the visibility of intra-cluster (b) and
The GIB layout can also be applied to gain further inter-cluster (c) edge structures. GIB provides clarity
insights into clustered graphs, in particular about the regarding the regular clique structures of the local
boundary spanners whose position within the original graphs neighborhoods within each cluster. It also enables the
can be obscured due to cluster occlusion. By generating identification of vertices with “long range” edges,
synthetic network datasets, with known structural properties, responsible for shrinking the network diameter.
we illustrate the conditions under which the GIB layout leads C. GIB Visualization of Pseudo-random Graphs
to new insights.
The pseudo-random graphs containing synthetic clusters
A. Datasets do not exhibit any particularly interesting intra-cluster
We used the CONTEST Matlab toolbox [22] to generate structure to analyze. However, they enable us to explore
the following synthetic network datasets: limitations regarding the visualization of clustered graphs for
- Small-world network with 500 vertices (avg. shortest- a range of inter-cluster edge probabilities. Additionally, they
path length = 10.349; clustering coefficient = 0.467), enable us to observe the positioning of boundary spanners
generated according to the Watts-Strogatz model [24]. both with and without GIB.
The visualizations in Figure 5 show two pseudo-random
- Two pseudo-random graphs, each containing 5 clusters networks with five clusters displayed using the HK fast
of different size (comprising 20, 40, 60, 80 and 100 multi-scale layout algorithm. GIB is used to isolate each of
vertices), generated according to the Erdös-Rényi model the five clusters into its own box. Figures 5 (a) and (b)
[6], with intra-cluster edge probability of 0.15. The two illustrate the effect of increasing the inter-edge density on the
graphs differ in the inter-cluster edge probability set to occlusion of the clusters and the re-positioning of boundary
0.005 and 0.05, respectively. spanners (i.e. the vertices with the largest size), which tend
to gravitate towards the periphery of the clusters, when these
- Scale-free network comprising 500 vertices, generated
are more separated. The GIB layouts, shown in Figures (c)
using the Barabási and Albert preferential attachment
and (d), display essentially the same intra-cluster structure
model [2]. Each new vertex was given 3 edges on arrival,
for both graphs. This is expected since both have the same
leading to a scale-free degree distribution.
intra-cluster edge probability of 0.15. In this case, GIB
enables a clear visualization of the centrality of boundary
B. GIB Visualization of a Small-world Network spanners within their own community.
Small-world networks are characterized by high clustering
coefficients and small average shortest path between pairs of D. GIB Visualization of a Scale-free Network
vertices. Vertices in such network are densely connected The scale-free network shown in Figure 6 (a) models
with only a small number of neighbors, but can be reached structural properties of real social networks with power-law
from any other vertex in the network through a short path. degree distributions. The CNM algorithm identified 10
The HK layout of the clustered small world network shown clusters in this network. The color of the vertices reflects the
in Figure 4 (a) while enabling the observation of some cluster membership while the size corresponds to the vertex
structural regularity, it obfuscates the details about the betweenness centrality. By removing the boundary edges and
structure of individual clusters detected by the CNM applying the HK layout algorithm we arrive at the graph (b),
algorithm. On the other hand, the GIB visualizations shown which renders separate clusters on top of each other.
in Figures 4 (b) and (c) isolate each of the five clusters into

(a) (b)

(c) (d)
Figure 5. Pseudo-random graphs with 5 clusters of different sizes (comprising 20, 40, 60, 80 and 100 vertices), with intra-cluster edge probability of 0.15:
(a) inter-cluster edge probability of 0.005; (b) inter-cluster edge probability of 0.05. The graphs are visualized using the Harel-Koren fast multi-scale layout
agorithm and vertices are sized by betweenness centrality. The visualizations in (c) and (d) are the corresponding GIB layout.

The GIB algorithm provides an improved and clearer boundaries, GIB adds value by showing the local, within-
view of the structure of individual clusters. It also shows that group link structure and inter-linking of the groups.
vertices with high betweenness play a central role within Our method of assessing the effectiveness of the GIB
individual clusters and are pulled towards the center of the layout is based on observations how well GIB reveals prior
clusters. This is expected in networks generated through knowledge about the networks. We also consider how well it
preferential attachments since such vertices also exhibit high abides by the design principle proposed by Krempel [17]:
closeness centrality. “constrain the solutions to simple patterns, patterns which
are known to put relatively little demand on the perceptual
VI. DISCUSSION AND RELATED WORK skills of an audience.” In that vein we considered relatively
standard notions and characteristics of networks that the
A. Observations from the GIB Experiments users are likely to observe and interpret by themselves.
Investigation of GIB across real and synthetic networks Experiments with the synthetic networks show that GIB
revealed several properties. First, GIB layout subsumes three identifies and confirms known properties of specific graphs:
important aspects in the visualization of networks: (1) it - The layout of small world networks highlights the long-
enables simultaneous filtering or grouping of vertices based range links between remote vertices. From inter-cluster
on a single or multiple attributes, (2) it enables layout links one can observe the cross-cluster effects of remote
optimization of individual groups while maintaining visual vertices (Figure 4).
awareness of other groups, and (3) it supports explicit - The layout of random networks shows boundary spanners
connection among close and remote vertices across groups. pulled centrally towards the direction of other clusters.
While some of these aspects can be achieved by simple Locally, within their own cluster, they are uniformly
filtering, the unifying effect that GIB offers is unique and not distributed, as expected. That is the case, both when the
easily achieved by other methods. links across clusters are sparse and when they are dense
Furthermore, with the focus on the local optimization of (Figure 5).
the group structure, GIB is robust and applicable across - In scale-free preferential attachment networks, the GIB
networks with different clusterability. Indeed, even in the local structure clearly highlights the presence of vertices
case of graph visualizations without clearly discerned cluster with both high closeness and betweenness centrality. This

(a) (b) (c)

(d) (e)
Figure 6. Scale-free network with 10 clusters detected by the Clauset-Newman-Moore algorithm. Vertices are colored by cluster membership and sized by
betweeness centrality. (a) Harel–Koren layout of the clustered graph. (b) Harel–Koren layout after removing inter-cluster edges. (c) Fruchterman-Reingold
layout after removing inter-cluster edges. (d) GIB showing inter-cluster edges and (e) GIB showing intra-clsuter edges.

is a defining property of such networks, giving advantage this sense, GIB expands the range of scenarios in which
to vertices and groups that form earlier in the network semantic substrates can be applied.
lifecycle (Figure 6). Harrer et al. [12] discuss various analyses of complex
Thus, for a range of analysis scenarios, from multi- networks and illustrate how rich semantic underpinning,
attribute characterization of networks to consistent insights such as ontologies that support user interaction around the
about network properties, GIB appears as a promising tool knowledge base, can lead to more complex network
for analyzing complex networks. representations. In such cases, one is likely to see a shift in
the network modeling, e.g., from a bi-partite graph analysis
B. Related Work on Multi-facted Network Analysis to multi-mode networks where more complex interactions
In their work on semantic substrates, Shneiderman and need to be expressed. In this view, the GIB layout covers a
Aris [20] promote two principles: (1) the use of semantic limited set of scenarios, where attributes themselves do not
substrates that are user-defined and (2) the use of controls, induce additional structure that needs to be taken into
such as sliders, to mediate the visibility and clutter in account. Moving towards a hierarchical representation of
visualization and ensure comprehensibility of the network categories would be the first step in that direction.
graphs. We extend this approach to semantic groupings of
nodes that can be expressed by a single or multiple attributes, C. Related Work on Layout of Clustered Graphs
as in the U.S. Senate voting analysis (Figures 2 and 3). We Clutter in network visualization may occur due
show that with GIB the user can explore semantic substrates positioning of the vertices or through interaction with the
that are based on either structural clusters or attribute-based graph, e.g., by exposing labels on vertices or edges. This
groupings of graph vertices. These, in turn, can be further issue has been addressed by modifying the original layout to
overlaid with additional attributes associated with vertices. In improve the readability. However, such transformations of

the graph layout need to be done cautiously to avoid [3] Clauset, A., Newman, M.E.J. and Moore, C. (2004) Finding com-
distorting the structural aspects of a network. Eades and Feng munity structure in very large networks. Physics Review E, 70, 06111.
[5], for example, address the issue of visualizing complex [4] Eades, P. (1984). A Heuristic for Graph Drawing. Congressus
Numerantium, 42:149-160.
clustered structures by providing multiple levels of
[5] Eades, P. and Feng, Q.-W. (1997). Multilevel Visualization of
abstraction that preserve the mental map of the original Clustered Graphs, Proc. Graph Drawing, LNCS, 1190: 101-112.
graph. In another approach, Henry et al. [13] propose using
[6] Erdös, P. and Rényi, A. (1959). On random graphs. Publicationes
vertex replication in order to emphasize the role of important Mathematicae (Debrecen), 6:290-297.
actors in the community and reduce inter-cluster edge [7] Girvan, M. and Newman, M.E.J. (2002). Community structure in
crossing [14]. However, while beneficial for a number of social and biological networks, Proc. National Academy of Sciences,
usage scenarios, this method does alter the network structure 99:7821-7826.
and may affect visual perception during analysis. [8] Freeman, L. C. (1977). A set of measures of centrality based on
betweenness.Sociometry, 40:35-41.
Other approaches take into account readability criteria by
[9] Fruchterman, T.M.J., and Reingold, E.M. (1991). Graph Drawing by
adding layout optimization constraints to the force-directed
Force-Directed Placement. Software: Practice & Experience, 21(11):
algorithms. For example, Kamps et al. [16] add a “vertex- 1129-1164.
vertex repulsion force" to the classical spring embedding [10] Hansen, D.L., Shneiderman, B., Smith, M.A. (2010). Analyzing
model [4] in order to minimize the overlapping area of two Social Media Networks with NodeXL: Insights from a Connected
vertices. Harel and Koren’s method [11] calculates a World. Morgan Kaufmann.
repulsive force by considering the variant of the distance [11] Harel, D. and Koren, Y. (2002). A Fast Multi-Scale Algorithm for
between the boundaries of vertex pairs. For clustered graphs, Drawing Large Graphs. J. of Graph Algorithms Appl., 6(3):179-202.
Noack [18] proposes optimization of planar layouts through [12] Harrer, A, Malzahn, N, Zeinin, S, and Hoppe, HU (2007). Combining
minimization of the LinLog energy which takes into account Social Network Analysis with Semantic Relations to Support the
Evolution of a Scientific Community, Proc. 7th Computer Supported
the clustering properties of the network. Collaborative Learning Conference, 267-276.
Within the multi-faceted network analysis, we apply a [13] Henry, N., Bezerianos, A. and Fekete, J-D. (2008). Improving the
hybrid approach, promoting GIB in the contexts where the Readability of Clustered Social Networks using Vertex Duplication.
network structure and category groupings need to be jointly IEEE Transactions on Visualization and Computer Graphics, 14(6):
analyzed. The criteria for combining the two will evolve 1317-1324.
from the investigation of usage scenarios and developed user [14] Henry, N., Fekete, J.-D. and McGuffin, M.J. (2007). NodeTrix: a
hybrid visualization of social networks. IEEE Transactions on
practices. Visualization and Computer Graphics, 13(6):1302–1309.
[15] Johnson, B. and Shneiderman, B. (1991). Treemaps: A space-filling
VII. CONCLUDING REMARKS AND FUTURE WORK approach to the visualization of hierarchical information structures,
In this paper we address the issue of supporting multi- Proc. 2nd International IEEE Visualization Conference, 284-291.
faceted analysis of networks through an extension of the [16] Kamps, T., Kleinz J. and Read, J. (1995). Constraint-Based Spring-
Model Algorithm for Graph Layout, Proc. Graph Drawing 95, LNCS
semantic substrate visualization approach. We introduce a 1027, 349-360.
novel graph layout algorithm, Group-in-a-Box (GIB), which [17] Krempel, L. (1993). Simple representations of complex networks:
enables simultaneous visualization of network clusters and strategies for visualizing network structure. Presented at the 3rd
categories of vertices in the network. European Conference for Network Analysis, Munich.
We demonstrate the benefits of the unified approach [18] Noack, A. (2008). An Energy Model for Visual Graph Clustering.
through experimentation with real and synthetic networks IEEE Transactions on Visualization and Computer Graphics, 14(6):
with diverse network properties. We show that the GIB 1333-1339.
layout is consistently effective in revealing the expected [19] Shneiderman, B. (1992) Tree visualization with tree-maps: 2-d space-
network structure in simulated data. filling approach. ACM Transactions on Graphics, 11(1):92-99.
The current design of GIB will be further refined and [20] Shneiderman, B., Aris, A. (2006). Network Visualization by Semantic
Substrates. IEEE Transactions on Visualization and Computer
expanded to support hierarchical clustering of networks and Graphics, 12(5): 733 - 740.
to incorporate support for taxonomy of categories. By [21] Smith, M. A., Shneiderman, B., Milic-Frayling, N., Mendes
observing how users apply GIB in different contexts, we Rodrigues, E., Barash, V., Dunne, C., Capone, T., Perer, A. and
expect to arrive at readability criteria for hybrid layouts like Gleave, E. (2009). Analyzing (social media) networks with NodeXL.
GIB. In multi-faceted analysis, the interconnection of two Proc. of the 4th International Conference on Communities and
distinct groups, based on categories and clusters, needs to be Technologies (C&T '09), 255-264.
supported in order to maximize the layout utility. [22] Taylor, A. and Higham, D.J. (2009). CONTEST: a controllable test
matrix toolbox for MATLAB. ACM Transactions on Mathematical
REFERENCES Software, 35(4), Article 26.
[23] Wakita, K. and Tsurumi, Y. (2007). Finding community structure in
[1] Asahi, T., Turo, D. and Shneiderman, B. (1995) Using treemaps to
mega-scale social networks, Proc. 16th International Conference on
visualize the analytic hierarchy process. Information Systems
World Wide Web, 1275-1276.
Research, 6(4): 357-375.
[24] Watts, D.J. and Strogatz, S.H. (1998). Collective dynamics of 'small-
[2] Barabási, A.L. and Albert, R. (1999). Emergence of scaling in
world' networks. Nature, 393:440-442.
random networks. Science, 286(5439): 509-512.

2011 IEEE Social Computing Nodexl: Group-In-A-Box

More Related Content

What's hot (19)

Similar to 2011 IEEE Social Computing Nodexl: Group-In-A-Box (20)

More from Marc Smith (20)

Recently uploaded (20)

2011 IEEE Social Computing Nodexl: Group-In-A-Box