J Vis

DOI 10.1007/s12650-016-0375-5

R E G UL A R P A P E R

Chenhui Li • George Baciu • Yunzhe Wang

Module-based visualization of large-scale graph

network data

Received: 30 November 2015 / Revised: 3 June 2016 / Accepted: 7 June 2016

Ó The Visualization Society of Japan 2016

Abstract The efficient visualization of dynamic network structures has become a dominant problem in
many big data applications, such as large network analytics, traffic management, resource allocation graphs,
logistics, social networks, and large document repositories. In this paper, we present a large-graph visu-
alization system called ModuleGraph. ModuleGraph is a scalable representation of graph structures by
treating a graph as a set of modules. The main objectives are: (1) to detect graph patterns in the visualization
of large-graph data, and (2) to emphasize the interconnecting structures to detect potential interactions
between local modules. Our first contribution is a hybrid modularity measure. This measure partitions the
cohesion of the graph at various levels of details. We aggregate clusters of nodes and edges into several
modules to reduce the overlap between graph components on a 2D display. Our second contribution is a
k-clustering method that can flexibly detect the local patterns or substructures in modules. Patterns of
modules are preserved by the ModuleGraph system to avoid information loss, while sub-graphs are clustered
as a single node. Our experiments show that this method can efficiently support large-scale social and spatial
network visualization.

Keywords Network visualization  Module grouping  Graph drawing  Information visualization 

Community detection

1 Introduction

Large network analysis and visualization have become an important challenge with the rapid increase in
network data, such as data generated from social networks, modern transportation networks, document
referencing, and academic citations. Graph connectivity patterns are of interest in network data analytics for
detecting unexpected connectivity and clustering features. This would allow us to discover and isolate points
of interest in massive networks which are otherwise difficult to spot visually on current displays. However,
the effect of large network visualization is often limited to the size of the display (Zinsmaier et al. 2012).
A practical approach to displaying a large-scale network is partitioning the network according to well-
defined domain-dependent attributes. However, graph visualization in the presence of incomplete infor-
mation is an open challenge and applications in this area can be found in abundance. To better visualize and

An earlier version of the paper was presented at the SIGGRAPH Asia 2015 Symposium on Visualization in High Performance
Computing, 2–5 November, Kobe, Japan, 2015.

C. Li  G. Baciu (&)  Y. Wang

Department of Computing, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, People’s Republic
of China
E-mail: [email protected]
 C. Li et al.

understand patterns in large-graph discovery, we focus on the representation of local patterns. This is a
critical step in deciding the structural components of a graph visualization.
In this paper, we present the ModuleGraph, a flexible graph visualization framework that aggregates the
components of large graphs into modules to help users to address graph-size limitations and effectively gain
insights into attributed graph data. The main idea of this paper is to partition the graph into several clusters
and visualize the relationships between clusters. By visualizing these clusters, we can find missing and/or
relevant associations. By increasing the level-of-detail, we can further emphasize graph patterns within
clusters and allow users to understand the structure of a large graph in detail.
We first address the problem of module-based graph visualization by introducing a graph modularity
measure. Second, a graph simplification method is adopted to accelerate the module detection. Third, each
graph pattern in the module is analyzed using the k-clustering method to enhance the module visualization.
Fourth, the visualization of ModuleGraph is constructed by (a) building a module-based graph that indicates
the basic structure of the original graph and (b) identifying the sub-graph patterns. We use a symbolic
signature instead of a simple node in the community graph. Finally, our experiments on real social networks
and spatial networks show that ModuleGraph can effectively transform large-graph data into recognizable
patterns and shapes that reveal significant structural and topological information.
The remainder of our paper is organized as follows. The related work is discussed in Sect. 2. We then
present the details of our ModuleGraph construction, including the mathematical model and the visual
design, in Sect. 3. We describe the experimental study to demonstrate the effectiveness of our method in
Sect. 4. In Sect. 5, we further discuss the performance of our methods. Finally, in Sect. 6, we conclude our
work and present future research directions.

2 Related work

The primary objective of large network visualization is the understanding of global and local patterns in
dynamic graphs, such as connectivity bundles, clustering, boundary formations, and expansions. Zinsmaier
et al. (2012) are inspired by the level-of-detail (LOD) techniques in the computer graphics discipline and
introduce a straight-line graph drawing that can be rendered interactively with different levels of detail to
visualize large-scale graphs. Because display systems have finite area and are physically limited both in size
and spatial dimensions for visualizing large-graph structures, a multi-screen solution is adopted in Chae
et al. (2012). However, the multi-screen approach imposes restrictions on spatial layouts and interactions.
Various methods for graph simplification have been developed. The most prominent ones are the graph
summarization (Tian et al. 2008), HiMap (Shi et al. 2009), and K-Core (Dorogovtsev et al. 2006). The
graph summarization method in Tian et al. (2008) allows the user to interactively control the resolution of
each aggregation in a large graph. HiMap (Shi et al. 2009) is presented to visualize large-scale social
networks through a hierarchical summarization. Complexity reduction via K-Core (Dorogovtsev et al. 2006)
is used to remove nodes with less than k links. Zhang et al. (2015) proposed a new model to calculate the
core structure with a million or more nodes. Unlike the k-core method, this model can convert low-degree
vertices into core nodes. For further graph simplification, a dimensionality reduction method is proposed in
van den Elzen et al. (2016). This method projects the abstract information into points in the two-dimen-
sional space.
Our work is related to the community detection algorithm. It is aimed at partitioning a large graph into
modules. Newman (2006) converts the problem into an optimization problem using a modularity-based
method called Modularity Classes. This method can measure the performance of the algorithm on the
community detection problem. Blondel et al. (2008) further present an acceleration method called Louvain
to reduce the computational complexity of the community detection problem. They demonstrate the per-
formance of Louvain on several large data sets. Moreover, dynamic graph evolution is achieved in Vehlow
et al. (2015) using community detection and a stream representation.
For the representation of community networks, Blondel et al. (2008) and Rosvall and Bergstrom (2008)
discuss in detail community aggregation. They aggregate the nodes of a community into a super-node. A
machine learning method, such as Belief Propagation (Horng et al. 2011), is adopted to explore large
graphs. In addition, Dunne and Shneiderman (2013) improve the graph visualization readability by drawing
different glyphs. Wu et al. (2015) utilize Voronoi maps to enhance the community visualization. Compared
with the previous methods cited above, our method focuses on designing a module for anticipating the visual
effects of the final visualization layout for large-scale networks.
Module-based visualization of large-scale graph network

Fig. 1 Simple example of module detection and ModuleGraph

3 Visualizing graph modules

We define the visualization of large-scale networks as a problem of module detection and clustering.
Figure 1 is a simple example of module detection. The input of our framework is the network data and the
corresponding number of edges and nodes. First, we aggregate the nodes into modules through modularity
measuring. Second, we abstract the features of detected modules and assign the modules with different
patterns using a k-clustering method. In this step, we design five patterns to approximatively represent the
structure of the module. Finally, the detected information is visualized according to the module design.
Our method, ModuleGraph, can be viewed as an abstraction of a large graph. The right part of Fig. 1 is
an example of ModuleGraph. The topology of a large graph is denoted by G ¼ ðV; EÞ, where V ¼
½vT0 ; vT1 ; :::; vTn ; ðvi 2 R2 Þ denotes the vertices and E denotes the undirected edges. The ModuleGraph is
defined as Mg ¼ ðM; EÞ, where M ¼ ½mT0 ; mT1 ; :::; mTn ; ðmi 2 R2 Þ denotes a set of modules and E denotes a
set of undirected edges, where each edge has a weight of w ¼ jEj. Each m is a vector with several features,
such as the module pattern, sub-node count, and sub-link count.
Overall, each module can be considered as a community in the graph. Hence, a community detection
method called Modularity Classes, as described in Newman (2006), can be adopted to find the communities
from a large graph. The method of community detection is better than the k-means clustering method,
because assignment of k is not required in advance. A further improved method for community detection,
called the Louvain, reduces the calculation of the modularity. This has been shown to be effective on large-
scale community detection by Blondel et al. (2008). The process of finding the modules consists of max-
imizing the modularity of the linked nodes. The modularity measure Newman (2006) can Q be defined as:
 
1 X ki kj
Q¼ Wij  kðsi ; sj Þ ð1Þ
2m i;j 2m

where W ij indicates the linking weight of node i and node j; ki and kj denote the counts of linked nodes on
node i and j, respectively; m indicates the sum of all linking weights; and the function kði; jÞ denotes whether
two nodes belong to the same community. The range of Wij is [0.0, 1.0]. Parameter kði; jÞ is set to 1 when
node i and j belong to the same community, or 0 when node i and node j belong to different communities.
Historically, researchers have paid less attention to community detection on spatial networks. An
example of a spatial network is a geographical network, such as airlines around the world. We further
consider the distance feature for a spatial network by modifying Eq. (1). The link with a longer length will
contribute less modularity. Hence, the improved modularity measure Qd can be defined as follows:
 
1 X a ki kj
Qd ¼ Wij distij  kðsi ; sj Þ: ð2Þ
2m i;j 2m

In Eq. (2), the value of a indicates the network type. The type could be 0 or 1, respectively, denoting a
non-spatial network or a spatial network. distij denotes the Euclidean distance of two nodes in a spatial
network. In a non-spatial network, the value of distij is 1.
The calculation of Modularity Classes (Newman 2006) for a large-scale graph is time consuming.
Although the Louvain method (Blondel et al. 2008) achieves a high performance, its precision is lower than
that of Modularity Classes. We present a hybrid modularity-based method to utilize the two addressed
methods. Our method is based on the modularity and is suitable for calculating the modules of a large-scale
graph. We can define the hybrid modularity Qh as follows:
 C. Li et al.

Fig. 2 Visualization of a simplified large network


Qd ; ðcv þ cl Þ  e
Qh ¼ ð3Þ
Qd 0 ; ðcv þ cl Þ [ e:
In Eq. (3), cv denotes the vertex count of the graph, cl denotes the edge count of the graph, and Qd 0 is a
high-speed community detection measure that removes the function of kðsi ; sj Þ. Although kðsi ; sj Þ removal
will slightly affect the accuracy of community detection, it makes sense of the community detection of a
large-scale graph. e 2 ð0; þ1Þ is a parameter that decides which measure will be used. The setting of e is
performed based on the computer configuration, namely the available CPU and RAM. In our experimental
environment, we set e as 0.1 million to achieve interactive visualization when the input network is a large-
scale network. If the experimental computer is more powerful than ours, as shown in Sect. 5, the user can
increase e to achieve more accurate results.
Using the hybrid modularity-based method, we can effectively detect the communities of a large-scale
graph, as shown in Fig. 2 (left). Different communities are assigned different colors. To observe the graph
clearly, each community can be replaced with a single node, as shown in Fig. 2 (right). A simplified
community graph allows the user to easily realize the clear structure of the original graph.
When the module sizes become very large, the simplified community graph may remain unsatisfactory
on a limited display. Our Module Detection method can be extended using the hierarchical module method.
A hierarchical module means that the detected communities in the previous detection process can be
considered as the input of a new community detection process.
For hierarchical module detection, we define the iteration of module detection as i, with a default value
of 1. Screen width and height are defined as w and h. If the number of nodes in the input network is far
greater than gwh, we increase the module iteration of i until the node count in the new graph is equal to or
less than gwh. g is a free parameter that reflects the blank space of the module visualization in the display. In
our experiment, we set g as 0.3, w as 1440, and h as 900.

3.1 Pattern definition

To further gain insight into modules, we analyze the module patterns. Wernicke (2006) presented a method
to detect motifs in a large network. Motifs indicate the link patterns in the graph. Unlike motifs, we calculate
several pattern factors to analyze each module type according to its internal graph, as well as other
significant features, such as module size, linking weight, visibility, and occupation percentage. An internal
graph represents the structure of the module.
To simplify the module representation, we classify the modules into five agent patterns. The design of
agent patterns is according to the most common network structures as presented in network structure theory
(Xu 2013) and communication theory (Theory 2010). Figure 3 shows our designed patterns: tree-connec-
tion, concentration, circular, low-connectivity, and full-connectivity. The tree-connection pattern indicates
that the graph is extended from a node and that some children of the root still contain the children. The
Module-based visualization of large-scale graph network

Fig. 3 Five types of graph modules

Fig. 4 Examples of graph modules: for each example, we show the agent pattern on the left, and example cases on the right

concentration pattern indicates that all children nodes are connected to the root and that each child does not
have other connections. The circular pattern indicates that each node in the graph only has two links. The
full-connectivity pattern is a complete graph. The remaining structure of the graph will be classified as a
low-connectivity pattern. We define the patterns in Fig. 3 as the agents to represent other similar structures.
We show some cases of similar structures of the agent patterns in Fig. 4. The right part of each sub-
figure shows some graph cases of the related patterns. The method of mapping the graph-to-agent patterns
will be discussed in Sect. 3.2.
 C. Li et al.

Fig. 5 Example of k-clustering. The color indicates the pattern type

3.2 Pattern detection via k-clustering

We present a k-clustering method to detect the pattern type of each module. To classify the module into the
agent patterns that we have defined, we abstract three features for each module. The feature vector can be
defined as vf ¼ ½a; b; d. We define the number of connected nodes for a node as c. By following the pattern
design, we can denote a ¼ maxðe P i Þ, b ¼ countðei ¼2Þ, and d ¼ 1 countðei [ nÞ, where ei denotes the neighbor
ei n n 2
count of node i and n denotes the node count of the module.
We propose a k-clustering method to approximately generate the clusters according to the module
feature vector. Each cluster includes the modules with the same pattern. The proposed clustering method
includes five steps. First, vf is calculated for each module. We define each module as a node in our method
and consider vf as a three-dimensional position. Initially, all the nodes are not assigned to any cluster.
Second, the cluster count (k) is initialized as 5, because we have designed 5 types of agent patterns for
visualization. Third, 5 nodes that match 5 agent patterns are selected, and their positions are assigned as the
cluster centers. The selection method can follow the agent pattern definition. Hence, we can obtain 5 initial
clusters. Each initial cluster only contains one node. Fourth, the remain nodes will be assigned to the closest
cluster by calculating the distance between its position and the other cluster center. Fifth, the mean position
of the cluster’s nodes is calculated and defined as the new cluster center.
The PfourthPand fifth steps should be repeated until the sum of the distances D is minimized. D is defined
i\k j\countðiÞ
as D ¼ i¼0 j¼0 distðpij ; ci Þ, where pij indicates node j in the cluster i, ci indicates the cluster center
of the cluster i, and dist is the distance calculation function. Figure 5 presents an example of 5 cluster
distributions in the three-dimensional space after the k-clustering process is completed. Colored circles
denote the cluster centers.

3.3 Visual design

The basic ModuleGraph layout can be calculated using a direct-force algorithm to ensure that each module
will not be overlapping with other modules. The main idea of our visual design for ModuleGraph is to
attempt to visualize the specific patterns instead of the crowd nodes in a large graph. The visual elements of
each module include a circle with a different radius called mc , an outside orbit with a different percentage
called mo , an icon called mp that indicates the module pattern, and a set of edges called me that connect the
related modules. Each edge between modules has a different width according to the weight. Figure 6 gives a
description of the ModuleGraph design.
We choose the color themes from the Google Color Palette (Google 0000) to represent different
modules. Because the color types are limited, the number in the center of the module can be used to indicate
the id of the module. When the module graph remains large scale after simplifying the original graph, we
design a chord diagram to assist in module visualization. A chord diagram is adopted to present the module
relationship and distribution through the circle layout according to the importance weight of the module.
Module-based visualization of large-scale graph network

Fig. 6 Visualizing the ModuleGraph design

Fig. 7 Chord diagram representation of ModuleGraph

The importance weight can be defined according to the neighbor count of a module. The outside circle’s
color indicates the pattern type. The importance percentage of a module and the relationships among
modules can be easily determined via a chord diagram, as shown in Fig. 7.

4 Experimental study

4.1 Social network

We select several large social network data sets from SNAP (Leskovec and Krevl 2014) which is a data
library for analyzing large information networks. The selected social network data set is from Facebook,
DBLP, YouTube, and Orkut, as summarized in Table 1. The visualization results of ModuleGraph for these
 C. Li et al.

Table 1 Selected social network data sets

Data sets Nodes Edges Module Brief modules

Facebook 4,039 88,234 16 –
DBLP 317,080 1,049,866 442 –
YouTube 1,134,890 2,987,624 9821 8
Orkut 3,072,441 117,185,083 1969 244

Fig. 8 Visualizing social networks by ModuleGraph and chord diagrams

network data are shown in Fig. 8. From the results, we can find that the module patterns can be directly
presented without any other interactions. From Fig. 8a, we can find that most of the module structure of the
Facebook network is a concentration structure. A high percentage of concentration modules shows that most
of the people on Facebook like to follow the hot people. Figure 8b is an assistant representation of Fig. 8a.
The outer circle in Fig. 8b shows the percentage of each module type. Each chord at the outside circle
represents a module. The inside part shows the connection of each module.
Figure 8c shows the ModuleGraph of DBLP data. The DBLP data include the paper citation relation-
ships of the researchers. Because there are many modules in the final result, we hide the outside circle of the
Module-based visualization of large-scale graph network

Fig. 9 Hierarchical ModuleGraph of YouTube social network

Fig. 10 Visualization of the size distribution of modules via box plot diagram for the Orkut network

ModuleGraph. From Fig. 8c, we can find that the module pattern appearing most often is the tree-connection
pattern. When the user clicks the module, as shown in the red rectangle of Fig. 8c, the related relationship of
the selected module can be presented through a chord diagram, as shown in Fig. 8d. The bottom part of
Fig. 8d presents the pattern types.
For large-scale networks, the hierarchical ModuleGraph representation as shown in Fig. 9 produces a
clearer visual representation which can aid significantly in querying and visual selections. The left side of
Fig. 9 shows the visual results without the hierarchical process. Because the detected modules are beyond
the screen size, we create a new ModuleGraph by making the previous ModuleGraph as the input graph. The
right side of Fig. 9 is the hierarchically processed result.
To further describe the information of the ModuleGraph, we adopt box plots to show the type distri-
butions of the modules, as shown in Fig. 10. Figure 10 is related to the module information of the Orkut
network. The five box plots in this Fig. 10 represent the five agent pattern statistics of the ModuleGraph.
Each box plot includes five values: maximum, second maximum, median, second minimum, and minimum.

4.2 Spatial network

For the case study of a spatial network, we select a data set with a massive number of flight records in the
US (Wickham 2011) in 2008. There are nearly 0.6 million flight lines, which are related to 3374 airports.
Figure 11a is the original visualization of the airline network. By comparison with the simple representation,
our method’s result as shown in Fig. 11b represents the airport community distribution and the bundling
airlines. In this case, the outside circle denotes the number of neighboring airports. The airline count inside
the module is presented via the size of the inside circle. The pattern icon for each module represents the sub-
network structure of the detected area. The information presented through ModuleGraph can help data
analyzers acquire the important patterns from a large-scale network. For example, the air traffic northwest of
New York State can be easily observed from Fig. 11b.
 C. Li et al.

Fig. 11 Spatial network of USA airline flights for a given period of time

Table 2 Comparison of module detection performance

Data sets Community detection Fast community detection (Blondel Module detection Module detection with
(Newman 2006) et al. 2008) (s) method (s) k-clustering (s)
Facebook 1.1 (s) 0.2 1.2 1.2
DBLP 217.4 (s) 11.0 12.5 13.1
YouTube Out of memory 25.3 27.4 27.7
Orkut Out of memory 247.1 251.5 251.6
Flight Lines 54.8 (s) 4.2 5.7 5.74

5 Implementation and evaluation

We use D3 (Bostock et al. 2011) to visualize the results on a computer with 2.8 GHz Intel i7 CPU and 16
GB RAM. The visualizations are displayed in the Chrome browser (45.0) with a 1440  900 resolution.
Module detection is implemented using the C?? language on the Windows 7 operating system.
Table 2 summarizes the performance of three module detection methods: community detection (New-
man 2006), fast community detection (Blondel et al. 2008), and our module detection method. We also
provide the module detection performance with k-clustering. From the summary, we can find that our
method can effectively support large-scale network aggregation and visualization. Because a distance
feature has been considered in our method, our method is slightly slower than the fast community detection
(Blondel et al. 2008) for large-scale network processing. The time cost of k-clustering is ignorable, because
it is only related to the number of detected modules.
Compared with the full network rendering, ModuleGraph achieves a satisfactory rendering performance
for basic interactions. Timing results are shown in Table 3 where fps indicates frames per second. The
rendering of the Orkut ModuleGraph is performed at high speed, because we adopted a hierarchical ren-
dering strategy for the very-large ModuleGraph. Although in some experiments, the massive data sizes
Module-based visualization of large-scale graph network

Table 3 Comparison of the rendering performance

Data sets Full network ModuleGraph (fps)

Facebook 0.19 (fps) 58.82
DBLP Out of memory 0.94
YouTube Out of memory 0.55
Orkut Out of memory 17.86
Flight Lines 0.15 (fps) 30.30

reduce the rendering speed, such as DBLP and YouTube, in Table 3, it is feasible to rewrite our Mod-
uleGraph rendering applications directly on a GPU subsystem to achieve interactive rates.

6 Conclusion and future work

The ModuleGraph is an effective visualization tool for aggregating and representing large-graph data
through clustering and interconnectivity pattern analysis. A module detection method is presented to
integrate the nodes that belong to the same group into a new integral module. We further enhance our graph
visualization method by representing the modules and the relationships among them. Our method can also
be applied to spatial network visualization. In the future, we plan to extend ModuleGraph to visualize
streaming networks. This would extend the ModuleGraph to a Dynamic ModuleGraph visualization system.
The context change between modules can also be considered. Moreover, our method can be extended to
visualize fuzzy networks that include overlapping communities.

Acknowledgments The authors would like to acknowledge the partial support of the Hong Kong Research Grants Council
Grants, GRF PolyU 5100/12E, IGRF PolyU 152142/15E, and Project 4-ZZFF from the Department of Computing, The Hong
Kong Polytechnic University.

References

Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E (2008) Fast unfolding of communities in large networks. J Stat Mech
Theory Exp 2008(10):P10008
Bostock M, Ogievetsky V, Heer J (2011) D3: data-driven documents. IEEE Trans Vis Comput Graph 17(12):2301–2309
Chae S, Majumder A, Gopi M (2012) HD-GraphViz: highly distributed graph visualization on tiled displays. In: Proceedings of
the eighth indian conference on computer vision, graphics and image processing, ICVGIP ’12, pp 43:1–43:8
Dorogovtsev SN, Goltsev AV, Mendes JFF (2006) k-core organization of complex networks. Phys Rev Lett 96:040601
Dunne C, Shneiderman B (2013) Motif simplification: Improving network visualization readability with fan, connector, and clique
glyphs. In: Proceedings of the SIGCHI conference on human factors in computing systems, CHI ’13. ACM, pp 3247–3256
Google: color palette. https://www.google.com/design/spec/style/color.html
Horng D, Chau P, Kittur A, Hong JI, Faloutsos C (2011) Apolo: making sense of large network data by combining rich user
interaction and machine learning. In: Proceedings of the SIGCHI conference on human factors in computing systems,
pp 167–176
Leskovec J, Krevl A (2014) SNAP Datasets: Stanford large network dataset collection. http://snap.stanford.edu/data
Newman MEJ (2006) Modularity and community structure in networks. Proc Natl Acad Sci 103(23):8577–8582
Rosvall M, Bergstrom CT (2008) Maps of random walks on complex networks reveal community structure. Proc Natl Acad Sci
105(4):1118–1123
Shi L, Cao N, Liu S, Qian W, Tan L, Wang G, Sun J, Lin CY (2009) Himap: adaptive visualization of large-scale online social
networks. In: 2009 IEEE pacific visualization symposium (PacificVis), pp 41–48
Theory C (2010) Patterns of communication. http://communicationtheory.org/patterns-of-communication
Tian Y, Hankins RA, Patel JM (2008) Efficient aggregation for graph summarization. In: Proceedings of the 2008 ACM
SIGMOD international conference on management of data, SIGMOD ’08, pp 567–580
van den Elzen S, Holten D, Blaas J, van Wijk J (2016) Reducing snapshots to points: a visual analytics approach to dynamic
network exploration. IEEE Trans Vis Comput Graph 22(1):1–10
Vehlow C, Beck F, Auwärter P, Weiskopf D (2015) Visualizing the evolution of communities in dynamic graphs. Comput
Graph Forum 34:277–288
Wernicke S (2006) Efficient detection of network motifs. IEEE/ACM Trans Comput Biol Bioinform 3(4):347–359
Wickham H (2011) Asa 2009 data expo. J Comput Graph Stat 20(2):281–283
Wu Y, Wu W, Yang S, Yan Y, Qu H (2015) Interactive visual summary of major communities in a large network. In: 2015
IEEE pacific visualization symposium (PacificVis), pp 47–54
Xu J (2013) Topological structure and analysis of interconnection networks, vol 7. Springer Science and Business Media
Zhang X, Martin T, Newman MEJ (2015) Identification of core-periphery structure in networks. Phys Rev E 91(3):1–10
Zinsmaier M, Brandes U, Deussen O, Strobelt H (2012) Interactive level-of-detail rendering of large graphs. IEEE Trans Vis
Comput Graph 18(12):2486–2495