Taurus: Towards a Unified Force Representation and

Universal Solver for Graph Layout

Mingliang Xue, Zhi Wang, Fahai Zhong, Yong Wang
Oliver Deussen, and Yunhai Wang

Abstract—Over the past few decades, a large number of graph layout techniques have been proposed for visualizing graphs from
various domains. In this paper, we present a general framework, Taurus, for unifying popular techniques such as the spring-electrical
model, stress model, and maxent-stress model. It is based on a unified force representation, which formulates most existing
techniques as a combination of quotient-based forces that combine power functions of graph-theoretical and Euclidean distances. This
representation enables us to compare the strengths and weaknesses of existing techniques, while facilitating the development of new
methods. Based on this, we propose a new balanced stress model (BSM) that is able to layout graphs in superior quality. In addition,
we introduce a universal augmented stochastic gradient descent (SGD) optimizer that efficiently finds proper solutions for all layout
techniques. To demonstrate the power of our framework, we conduct a comprehensive evaluation of existing techniques on a large
number of synthetic and real graphs. We release an open-source package, which facilitates easy comparison of different graph layout
methods for any graph input as well as effectively creating customized graph layout techniques.
Index Terms—Graph Layout, Gradient Descent, Framework

1 I NTRODUCTION
Graphs are commonly used for modeling complex data in many To alleviate the involved computational costs, sparse stress models [13,
domains such as social media, finance and biology. The most commonly 27, 32] have been proposed, which only impose springs for a subset of
used graph visualization technique, node-link diagrams, depict nodes node pairs.
as points in a plane and edges as lines connecting these points. In past Because of the divergent mechanisms, these models have different
decades, various graph layout methods [25, 36] have been developed characteristics when creating graph layouts. For example, FDP
for producing aesthetically-pleasing drawings, while maintaining the performs better in preserving neighborhood structures for many
underlying graph structures. graphs, while the stress model tends to maintain the overall structures,
Rather than directly optimizing aesthetic criteria [34] (e.g., even especially for mesh-like graphs. However, it is still unclear why the
node distribution and minimal edge crossing), most methods simulate models have such differences and how they are connected conceptually.
one of two kinds of physical systems as a basis for layouting graphs: Moreover, it is difficult to make a fair quantitative comparison because
the spring-electrical model or the stress model. The spring-electrical different optimization strategies are used. This not only hinders
model [8, 10] regards edges as springs that use attractive forces to pull researchers to develop new methods but also poses a challenge for
connected nodes close to each other, at the same time treating nodes as practitioners to choose a proper method for visualizing their graphs.
electrically-charged particles that repel each other with repulsive forces. In this paper, we present a general framework, we call Taurus,
Based on this model, many variants of force-directed placement (FDP) Towards a Unified force Representation and Universal Solver for graph
algorithms have been developed for better revealing different structures layout, that offers a unified view for understanding and comparing
and features of graphs. For example, FM3 [18] and SFDP [20] use a most of the popular graph layout algorithms. It relies on two novel
multilevel scheme for overcoming local minima, the extended models components: a unified force representation and a universal solver.
of LinLog [31] and ForceAtlas2 [22] allow to better reveal clusters The uniform force representation allows us to show that all existing
and local structures, respectively. While the spring-electrical model methods can be formulated as a combination of quotient-based forces,
produces good layouts for many graphs, it does not encode the target using a quotient between power functions of graph-theoretical and
(data-space) edge lengths between every pair of nodes. Euclidian distances. This unified representation enables us to compare
This is the focus of stress models [14, 23, 40], which assume a spring the strengths and weaknesses of different methods. The universal solver
between each pair of nodes with an ideal length equal to the graph- combines the advantage of SGD [30] in escaping local minima and
theoretical distance among the nodes. By minimizing the stress energy the effectiveness of the Barnes & Hut approximation [3] in reducing
of the spring system, a layout is obtained. For efficiently solving such computational cost, which allows us to solve different existing layout
models, which involve considerably more interactions between the methods with the same optimizer.
nodes, a few optimization strategies have been incorporated, such as Moreover, our framework can also be used as a general platform
stress majorization [14], and stochastic gradient descent (SGD) [42]. for developing new graph layout methods. In particular, we propose
a balanced stress model, which combines the advantages of spring-
electrical and stress models. Specifically, it exerts attractive and
• Mingliang Xue, Zhi Wang, Fahai Zhong, and Yunhai Wang are with the repulsive forces to all node pairs, where the attractive force is reciprocal
Department of Computer Science, Shandong University, China. E-mail:
and the repulsive force is proportional to the graph-theoretical distances.
{xml95007, wangzizi2020, zhongfahai, cloudseawang}@gmail.com
In doing so, the model avoids extremely large repulsive and attractive
• Yong Wang is with School of Computing and Information Systems, Singapore
Management University, Singapore. E-mail: [email protected]
forces for nearby nodes, while pulling neighboring nodes close to each
• Oliver Deussen is with Computer and Information Science, University of other.
Konstanz, Konstanz, Germany. E-mail: [email protected] We implement Taurus as a graph visualization package in C++,
• Yunhai Wang is the corresponding author which allows users to define their own attractive and repulsive forces.
To demonstrate its effectiveness, we comprehensively evaluate it by
Manuscript received xx xxx. 201x; accepted xx xxx. 201x. Date of Publication comparing various spring and stress layout methods on a large number
xx xxx. 201x; date of current version xx xxx. 201x. For information on
of synthesized graphs with different structures such as lattices, trees
obtaining reprints of this article, please send e-mail to: [email protected].
and clusters. The evaluation includes two parts: verifying whether
Digital Object Identifier: xx.xxxx/TVCG.201x.xxxxxxx
Taurus can produce similar results to the original implementations of
existing methods, and examining how different methods behave on Maxent-stress model (Maxent) [13] imposes stress constraints on pairs
graphs with different characteristics. The results show that our solver of neighboring nodes and entropy-based constraints on the remaining
enables all methods to perform as well as or even better than the original node pairs, the latter ones can be regarded as repulsive forces between
implementations, while our proposed balanced stress model makes a all node pairs.
good trade-off in distance preservation and maintaining neighborhoods Noack [31] shows that energy-based layout methods like LinLog can
as well as cluster structures. In addition, we show that our Taurus be formulated as force representations. Similarly, Gansner et al. [13]
allows users to flexibly customize the graph layout methods for meeting represent the repulsive force as an entropy term and incorporate it into
specific requirements. the stress-based energy model. However, there is still a lack of an
The main contributions of this paper can be summarized as follows: inherent representation for unifying existing layout methods. In this
work, we demonstrate that almost all methods from spring-electrical
• We propose a general framework for graph visualization based and stress models can be formulated as a combination of our proposed
on a novel quotient based force representation and an augmented quotient-based forces. Moreover, we show that this unified view not
SGD optimizer, which offers a unified view for understanding only facilitates the understanding and comparison of different methods
and comparing existing graph layout methods; but also allows the development of new methods.
• We present a new graph layout method based on our framework 2.2 Graph Layout Solvers
and conduct a systematic analysis and extensive evaluation for Most graph layout methods need an optimization solver to create
our framework on different graph datasets through quantitative desirable drawings. Solving a spring-electrical model has a time
comparisons; and complexity of O(n2 ) at each iteration, where n is the number of
nodes in the graph. To improve the computational efficiency of such
• We release a library with the proposed general framework that models, several multilevel methods [3, 11, 12, 18–20, 39] have been
enables rapid implementation and design of graph layout methods proposed. Among them, the Barnes-Hut (BH) approximation [3] is
for any graph input. the most commonly-used acceleration method. It uses hyper nodes
2 R ELATED W ORK to approximate repulsive forces, resulting in a time complexity of
O(n log n). The method has been used by different spring-electric
Related works can be categorized into three parts: graph layout model algorithms, such as [20, 22]. Another method is to use random
methods, graph layout solvers and graph layout packages. vertex sampling (RVS) [17] to accelerate the computation of repulsive
2.1 Graph Layout forces. This method generates layouts similar to Barnes-Hut.
There are also many algorithms to optimize solutions for the stress
Various graph layout methods have been proposed to visualize network
model. The earliest stress model [23] employs gradient descent to
data as node-link diagrams. Among them, the most common methods
find the optimal graph layout; however, it is often trapped into a local
often use virtual physical models to represent the relationships between
minimum. Gansner et al. [14] adapt stress majorization to the stress
objects. By referring to the taxonomy by Gansner et al. [13], we classify
model, which is rooted in solving multidimensional scaling. Ensuring
such methods into three types: spring-electrical models, stress models
a monotonic decrease of the stress, the method has advantages over the
and hybrid models.
original implementation. Recently, stochastic gradient descent (SGD),
Spring-electrical models [8, 10] regard nodes as electrically-charged
a powerful optimization solver widely used in machine learning, has
particles that push nodes away from each other and edges as springs
also been applied to graph drawing [42]. It converges fast and achieves
that pull nodes close to each other, often referred to as repulsive and
layouts with a lower stress error. Ahmed et al. [1] further proposed a
attractive forces. A graph layout result is achieved when attractive
SGD-based graph drawing approach (SGD)2 that can handle multiple
and repulsive forces strike a balance. For a complete review of the
readability criteria of graph drawing simultaneously. We propose
graph layout methods developed from this model, please refer to
an augmented SGD solver for finding optimal layouts at minimal
Kobourov [28] and Gouvêa et al. [16]. Here, we briefly review some
computation speed.
widely used models. Hu et al. [20] improve the repulsive force designed
by Fruchterman and Reingold [10] and use a repulsive force that decays 2.3 Graph Layout Packages
rapidly, avoiding edge-length distortion at the periphery of a layout. A number of open-source packages facilitate an easy implementation
Noack et al. [31] introduce the LinLog model that employed a constant of different graph layout techniques. For example, Graphviz [15],
attractive force and set the repulsive force to the inverse of the distance. Tulip [2] and OGDF [6] are C++ libraries that implement customized
As a result, this model can generate graph layouts with clearly-separated graph data structures and many graph drawing techniques. Data-Driven
node clusters. Kermarrec and Moin [26] further extend the LinLog Documentation (D3) [4], the most popular web-based visualization
model for revealing cluster structures at different levels. Inspired by toolkit, incorporates some graph drawing techniques (e.g., the spring-
these studies, the attractive and repulsive force of ForceAtlas2 [22] were electrical model [10]). All packages allow users to directly use
designed to be proportional and inversely proportional to the distance different graph layout methods without implementing them from
between nodes, obtaining graph layouts with a good preservation of scratch. Because of the underlying models, however, these packages
local structures and cluster separation. often expose different APIs and parameters for different methods,
Stress models [14, 23] also use a spring analogy but assume that there resulting in cumbersome parameter tuning for the user and the need for
are springs connecting every pair of nodes in the graph. Spring forces understanding different approaches. Building upon our unified force
are defined to create a layout with distances of nodes as close as representation and universal solver, our graph drawing package is much
possible to the graph-theoretical distances. Many variants of the stress more generic and easier to use. Different solutions can be compared
model aim to improve its efficiency through sparse approximations. and the right method for the wanted layout can be selected.
For example, progressive multidimensional scaling [5] and low-rank
stress majorization [27] have been used to approximate the shortest 3 P ROPOSED F RAMEWORK
path distances of all node pairs in a graph. The sparse stress model [32] As mentioned above, our general framework aims to unify existing
speeds up the stress model by aggregating the terms of the objective graph layout methods. It consists of a quotient-based force model to
function. Wang et al. [40] improved the stress model by imposing describe the relationship among nodes, and a universal optimization
constraints on edge vectors and edge lengths, further enhancing the solver to achieve optimal graph layouts. In this section, we first show
expressiveness of the stress model. how the proposed framework originates from the observations of prior
Hybrid models combine both models for overcoming their drawbacks. graph layout approaches. Then, we present our quotient-based force
For example, Hu and Koren [21] resolve the warping effect of spring- model as well as the guidelines for using it. Finally, we introduce our
electrical models by integrating attractive forces into the stress model. proposed balanced stress model.
To reduce the cost for computing graph-theoretical distances, the
 Table 1. Quotient based force functions and their corresponding
3.1 Revisiting Existing Graph Layout Methods
parameters of different layout methods: ω is the weight, α and β are the
For a graph G(V, E) with V 2 representing the set of node pairs, graph exponents of the graph-theoretical distance and the Euclidean distance
layout methods aim to map the graph nodes V to coordinates in 2D between two nodes, respectively, and Ω is the force range. P is a set of
or 3D space and often require a model to represent the relationship pivot nodes [27], k f a is defined as −(deg(i) + 1)(deg( j) + 1) [22] with the
between them. Depending on the underlying mechanism of building node degree deg(i). V 2 refers to all node pairs, E to node pairs connected
the model, Hu et al. [13, 21] classified layout methods into two types: by an edge, S to a k-ring neighborhood graph.
spring-electrical models and stress models. They propose to use
hybrid models, which integrate spring-electrical and stress models. Method Attractive Force {ω1 , α1 , β1 , Ω1 } Repulsive Forces {ω2 , α2 , β2 , Ω2 }
Spring-electrical models often use force modeling, while stress and FDP ∑(i, j)∈E ||xi − x j ||2 ei j {1, 2, 0, E} ∑{i, j}∈V 2 ||xi−1
−x j || ei j {-1,-1,0,V 2 }
hybrid models are built on energy modeling to specify the graph layout. [10]
Since the force on an object is the negative derivative of the energy kfa
{k f a , −1, 0,V 2 }
FA2 ∑(i, j)∈E ||xi − x j ||ei j {1, 2, 0, E} ∑{i, j}∈V 2 ||xi −x j || ei j
with respect to the distance [41], we re-write all energy-based layout [22]
methods into the form of a force modeling for establishing a unified −1
representation. In the following, we take one representative method of LinLog ∑(i, j)∈E 1 ∗ ei j {1, 1, 0, E} ∑{i, j}∈V 2 ||xi −x j || ei j {−1, −1, 0,V 2 }
[31]
each model type as an example.
2||xi −x j || −2
Force-Directed Placement. As a typical instance of the spring- SM [14] ∑{i, j}∈V 2
di2j
ei j {2, 1, 2,V 2 } ∑{i, j}∈V 2 di j ei j {−2, 0, 1,V 2 }
electrical model, FDP [10] aims to meet the principles that connected 2||xi −x j ||
nodes should be drawn near each other and all nodes should not be MARS ∑(i, j)∈P×V di j ei j {2, 1, 1, P ×V } ∑(i, j)∈P×V −2ei j {−2, 0, 0, P ×
[27] V}
drawn too close to each other. It computes the position of each node xi
2||xi −x j ||
by exerting the attractive force Fi,aj and repulsive forces Fi,r j between SSM ∑(i, j)∈P×V di2j
ei j {2, 1, 2, P × ∑(i, j)∈P×V −2
di j ei j {−2, 0, 1, P ×
∪E ∪E
the node and its neighbours and all other nodes, respectively. [32] V ∪ E} V ∪ E}
2||xi −x j || −2
Maxent ∑{i, j}∈S ei j {2, 1, 2, S} (∑{i, j}∈S di j + {−2, 0, 1, S},
x j − xi di2j
{−αsgn(q),
ei, j = , (1) [13] −αsgn(q)
∑{i, j}∈V 2 ||xi −x j ||q )ei j −q,0,V 2 }
||xi − x j ||
Fi,aj = ||xi − x j ||2 ∗ ei, j , ∀{i, j} ∈ E, (2)
1 Maxent-Stress Model. Instead of specifying springs for all node pairs,
Fi,r j = − ∗ ei, j , ∀{i, j} ∈ V 2 , (3)
||xi − x j || the maxent-stress model [13] is a hybrid model that defines a stress
model constraint on a subset of node pairs (typically, the set of graph
where ei, j is a unit vector. By successively moving each node along the edges E), while imposing an entropy-based constraint to the rest of the
resultant force Fi , node pairs. Hence, the energy function is defined as follows:

Fi = Fi,aj + Fi,r j , (||xi − x j || − di j )2


∑ ∑ sgn(q)
{i, j}∈E {i, j}∈V 2

 ∑ +α ∑ q
, if q 6= 0
di2j ||x i − x j ||


 (i, j)∈S (i, j)∈S
/
U=
the final layout is obtained when the force system reaches an  (||xi − x j || − di j )2
+ α ∑ ln ||xi − x j ||, if q = 0

equilibrium. To meet given layout principles, the unit vector ei, j in 
∑ di2j


Eqs. 2 and 3 can also be computed in terms of some constraints [7] (i, j)∈S (i, j)∈S
/
(e.g., node non-overlapping and minimal edge crossing).
Stress Model. Unlike the spring-electric model, this model aims to where the default S is E but can also be the k-neighborhood graph,
preserve predefined edge lengths in the visualization. It assumes that α > 0 and q > −2. When q is not zero, the gradient of this model is:
there is a spring between every pair of nodes with an ideal spring length,
which is equal to their graph-theoretical distance being one unit. Hence, ∂U 2(||xi − x j || − di j ) sgn(q)

=− ∑ 2
−α ∑ q+1
∗ ei, j ,
it obtains an optimal layout by minimizing the energy function: ∂ xi (i, j)∈S d ij (i, j)∈S
/
q||xi − x j ||

(||xi − x j || − di j )2 otherwise the gradient is:

U= ∑ , (4)
{i, j}∈V 2
di2j
2(||xi − x j || − di j )
 
∂U 1
=− ∑ − α ∑ ∗ ei, j .
where di j denotes the graph-theoretical distance between nodes i and ∂ xi (i, j)∈S di2j (i, j)∈S
/
||xi − x j ||
j. The original model is solved by using the gradient descent method
with the gradient: We can see that its attractive force has the same form but is only exerted
to a subset of node pairs, while two repulsive forces are exerted in
∂U 2(||xi − x j || − di j ) different ranges (as defined by S)
=− ∑ ∗ ei, j .
∂ xi
{i, j}∈V 2 di2j
2||xi − x j || 2
Fi,aj∈S = ∗ ei, j , r1
F(i, j)∈S = − ∗ ei, j , (6)
Since the negative gradient of the energy here is regarded as the acting di2j di j
force, we write Eq. 4 as forms of attractive and reclusive forces: r2 sgn(q)
F(i, / = −α
j)∈S ∑ q||x − x j ||q+1
∗ ei, j . (7)
2||xi − x j || 2 (i, j)∈S
/ i
Fi,aj∈V 2 = ∗ ei, j , Fi,r j∈V 2 =− ∗ ei, j . (5)
di2j di j
When q = 0, the second repulsive force is:
For escaping local minima, stress majorization [14], a widely used 1
r2
method for MDS solutions, has been adapted for solving this model. F(i, / = −α
j)∈S ∑ ||x − x j ||
∗ ei, j . (8)
(i, j)∈S i
To reduce the computation cost, a few extended stress models choose a /
subset of node pairs to compute the stress energy. For example, the low-
rank based stress model, Mars [27] and sparse stress model (SSM) [32] Besides these three methods, Table 1 lists the attractive and repulsive
both define the force range based on a set of pivot nodes, while SSM forces of a few other methods such as FA2 [22], Linlog [31], and
further incorporates the edge information. MARS [27].
 3. If αβ is larger than zero, the factors di j and ||xi − x j || have the
opposite effect on the force magnitude.

Case 1 corresponds to the forces defined in FDP (see Eqs. 2 and 3)

and the repulsive forces used in the stress model (see Eq. 5). In contrast,
there is only one example (see the attractive force in Eq. 5) for Case 3
in Table 1 and no example for Case 2. In the following, we provide the
guidelines for selecting proper α and β .
3.3 Guidelines for the Selection of Exponents
To faithfully maintain the relationship between nodes, we generalize a
core principle of the FDP method about connected nodes to all node
pairs. In FDP connected nodes should always be closer to each other
than to other nodes.

• Nodes with small graph-theoretical distances should be drawn

closer to each other than nodes with large distances.

Fig. 1. (a) Influence of the parameters {α, β } on the force. Each plot For simplicity, we divide all k forces within the graph into attractive and
shows the force magnitude as a function of the graph-theoretical distance repulsive forces. To meet the above principle, two nodes with a larger
between two nodes in the graph and the pairwise Euclidian distance graph-theoretical distance should be exerted a larger repulsive force
in the layout for the given combination of α and β . The yellow color (β < 0) and a smaller attractive force (β > 0). To prevent the layout
represents a force magnitude close to zero and the orange color a large from diverging to infinity or collapsing into a point, for two nodes with
force magnitude. The red and blue boxes cover the parameter settings fixed graph-theoretical distances, the repulsive force should decrease
satisfying the criteria G1 and G2, respectively. (b) shows the resultant as the Euclidian distance between the two nodes increases (α < 0), and
forces for FDP, (c) for the stress model and (d) for the maxent stress the attractive force should decrease as the Euclidian distance between
model. two nodes decreases (α > 0). For yielding a clustering (dispersing)
3.2 Quotient based Force Function effect, we can also use a constant repulsive (attractive) force with a
After systematically comparing and analyzing various forces used in large attractive (repulsive) force by setting α = 0 (β = 0). Therefore,
different graph layout methods (see Table 1), in the following we we identify the following two guidelines for choosing α and β :
identify common components that appear in most methods, and further
propose a quotient-based representation to unify them. Given l forces, • G1: For the attractive force, the exponent parameters are
the resultant force Fi exerted on a node i is: suggested to satisfy: α ≥ 0, β ≥ 0; and

l • G2: For the repulsive force, the exponent parameters are

||xi − x j ||αk suggested to satisfy: α ≤ 0, β ≤ 0.
Fi,k = ∑ ωk ∗ β
∗ ek,i j , Fi = ∑ Fi,k (9)
(i, j)∈Ωk di jk k
The examples in Fig. 1 enclosed by the red box correspond to
where Ωk is the force range specifying the nodes that exert forces parameters meeting G1 and blue box to parameters meeting G2.
towards node i, ωk is a weight with a sign deciding the force type Looking again at Table 1, we see that the attractive forces of all
(attractive force vs. repulsive force), αk is the exponent of the graph- methods satisfy G1, whereas the repulsive force of some existing
theoretical distance, βk of the Euclidean distance between nodes j and i, graph layout methods violate G2. For example, the repulsive forces of
which decide the force magnitude. To illustrate how this representation SM [14], MARS [27] and SSM [32] is −2 di j are not in the blue box, but in
unifies existing graph layout methods, we formulate some popular the red box of Fig. 1. Since their magnitude decreases with increasing
methods in this representation, see Table 1. Taking the maxent-stress graph-theoretical distance, the repulsive force might not be able to repel
model as an example, its attractive and first repulsive force have the node pairs with large graph-theoretical distances far from each other,
same force range {i, j} ∈ S, and their corresponding {ω, α, β } are resulting in false neighborhoods.
{2, 1, 2} and {−2, 0, 1}, respectively. The range of the other repulsive Reflections. To explore how different choices of α and β influence
force Ω is {i, j} ∈
/ S, the parameters {ω, α, β } are {−q sgn(q), −q, 0}. the final layout, we compute the resultant force for the two nodes with
Note that low-rank stress majorization (MARS) [5] and the sparse stress varying di j and Euclidean distance ||xi − x j ||. Figs. 1 (b,c,d) show the
model (SSM) [32] exert forces for ranges defined by the user specified results for three methods: FDP, stress model and maxent-stress model,
pivot nodes P. where a positive value indicates that the attractive force is larger than
Parameter Effect. A closer look at the examples in Table 1 tells us the repulsive force and vice versa for a negative value.
that there should be at least two forces (l ≥ 2) with different signs In Fig. 1(b), we can see that FDP exerts large repulsive forces on
for weights ωk to ensure convergence. Common force ranges are E, nodes with small Euclidean distance, but only applies attractive forces
V 2 or the ones defined by the user-specified node sets P. In contrast, to connected nodes. In other words, it treats all nodes with di j being
there are many choices of {αk , βk }, each of them might create different larger than 1 equally, resulting in a poor overall distance preservation.
behaviors. In the following, we show how existing methods set these Fig. 1(c) allows two observations about SM: i) the resultant force is
parameters. close to zero in most places (see the yellow region), except the ones for
To investigate the effect of {αk , βk }, we depict the force magnitude nodes with small graph-theoretical distances; and ii) the force exerted
between two nodes i and j as a function of their pairwise Euclidean on nearby nodes with large graph-theoretical distances is close to zero
distance ||xi − x j || and the graph-theoretical distance di j . Fig. 1(a) (see top-left corner). In other words, the original stress model might
shows the force magnitudes under different combinations of a few not be able to efficiently preserve local neighborhood structures and
commonly used α and β values {−1, 0, 1}. From the plots, we can see large graph-theoretical distances.
that there are three kinds of the relationship: For the maxent-stress model, we set S and q to E and zero and show
the resultant forces for two nodes in Fig. 1(d). When di j is larger than
1. If αβ is zero, the force magnitude is purely determined by ||xi −
one, its resultant force is equal to the one of FDP, as shown in Fig. 1(b),
x j || or di j ;
otherwise it is the same as the one of the stress model shown in Fig. 1(c).
2. If αβ is smaller than zero , the factors di j and ||xi − x j || have the Hence, we conclude that this model can be regarded as the combination
same positive or negative effect on the force magnitude; and of the spring-electrical and the stress model.
3.4 A Balanced Stress Model (BSM) 4 AUGMENTED SGD S OLVER
The above considerations and the possibility to formulate different Before we evaluate the balanced stress model in Section 5, we outline
methods with a unified formula help us to formulate a balanced stress our second contribution that allows us to compare all the mentioned
model that would have an “ideal” behavior. It is surprisingly simple models. Recently, Zheng et al. [42] adapt the stochastic gradient descent
and in contrast to many methods we would consider to be “ad hoc” in technique, a powerful optimization solver widely used for training deep
their selection of weights and exponents its behavior is directly derived neural networks, to solve the stress model. They show that SGD can
from the nature of the underlying problem. reach lower stress errors faster than stress majorization [14], while
As mentioned in Section 3.3, it is desirable to exert large repulsive not requiring a good initialization. However, the method is originally
force to the node pairs with large graph-theoretical distances di j , a fact designed for minimizing energy functions and thus has not been used
that the stress model does not satisfy. To address this issue, we propose for solving force-based layout methods.
the following model: Since energy is the negative integral of a force, we can formulate
each of the quotient-based force models (see Eq. 9) as the following
||xi − x j || di j energy:
Fi = ∑ ( − ) ∗ ei j , (10)
2 di j ||xi − x j ||
i, j∈V
||xi − x j ||αk +1

ωk

 ∑ ∗ αk 6= −1
 (i, j)∈Ω αk + 1 β

where the second term can be taken as a weighted repulsive force 
k d k ij
from the graph-theoretical distance as used in FDP (see Eq. 3), its Uk = (11)
reciprocal corresponds to the term of the attractive force. Doing so,
 ln ||xi − x j ||
otherwise.

nodes with large graph-theoretical distances are repelled far from each

 ∑ ωk β
di jk

(i, j)∈Ωk
other. On the other hand, a model after Eq. (10) is able to effectively
preserve graph-theoretical distances, since the resultant force is zero To minimize this energy, SGD repeatedly randomly picks a pair of
when the graph-theoretical distance between two nodes is equal to their nodes xi and x j for moving it along the force direction at a time:
Euclidean distance. To our understanding, this is a meaningful setup
that builds on the intuition that a layout should reflect graph-theoretical ωk ||xi − x j ||αk
distances as good as possible in its projected distances. r= ∗ ∗ ek,i j (12)
2 β
d k
As shown in Fig. 2 (a), the resultant forces exerted on two nodes ij
forms a skew-symmetric matrix and all forces are zero at the diagonal xi = xi − ηs r, x j = x j + ηs r, (13)
of the matrix. For us, an ideal behavior.
where ηs is the step size. With a carefully chosen step size, this method
quickly converges to a reasonable layout. Unlike Zheng et al. [42],
our SGD solver updates each node with two steps, since the gradient
of the stress model in Eq. 4 is separated into attractive and repulsive
forces. This facilitates us to apply SGD to solve force-based methods.
However, directly applying SGD to FDP-based methods might be too

Algorithm 1 Pseudocode for our augmented SGD solver

1: Input: graph G = (V, E),
2: X = RandomMatrix(|V |, 2)
3: for i = 1 to iters do
4: for k = 1 to l do
5: if ωk < 0 and βk = 0 then
6: solve_BH_forces(X, αk , βk , Ωk )
7: else
8: solve_SGD_forces(X, αk , βk , Ωk )
9: end if
10: end for
11: end for
Fig. 2. (a) Resultant force for two nodes with varying graph-theoretical 12: return X
and Euclidian distances generated by our balanced SM. (b) Layouts
generated by SM and balanced SM (c) on the graph cluster_2000.
Our method clearly separates different clusters. (d,e) The boxplots expensive because of its O(n2 ) time complexity. To alleviate this
summarize the value ranges (same for all boxplots in this paper) of the issue, we suggest to combine it with the Barnes-Hut (BH) [3] method,
stress error (d) in terms of different graph distances and neighborhood which approximates the repulsive force between nodes by using a
preservation (e) degrees for different ring sizes. quadtree structure with a time complexity of O(n log n). Note that
this approximation only holds for forces with parameters ωk < 0 and
Figs. 2(b,c) compare SM with our balanced model BSM using βk = 0. With this augmented SGD(ASGD) solver, we first compute the
cluster_2000. BSM clearly reveals the cluster structures, while SM repulsive force using the BH solver to move nodes and then use the SGD
shows much less separation. To further investigate the differences solver to update nodes at each iteration, as outlined in Algorithm 1.
between the methods, we calculate the stress error for different graph- Fig. 3(a) shows the convergence curves of different solvers for
theoretical distances and the neighborhood preservation degrees of computing the FDP layout of the btree9 graph. Our solver performs
different ring sizes. As shown in Figs. 2(d,e), BSM performs slightly similarly to the SGD solver, while both converge to smaller energy
worse than SM for smaller graph-theoretical distances and similar values than the BH solver. On the other hand, our solver is faster at
or even better for large distances, whereas it outperforms SM in each iteration and its overall time is lower than the SGD solver (see
neighborhood preservation no matter what the ring size is. These Section 5). Moreover, it is surprising that our solver even produces
results are consistent with our reflections about SM in Section 3.3 and better layouts than the SGD solver. As shown in Figs. 3(b,c,d), only our
hence we speculate that our BSM performs better in balancing distance solver is able to maintain the tree structure of the data. We speculate
preservation and neighborhood preservation. that this is due to the separation of attractive and repulsive forces for
node movement.
 • Clustered graphs refer to networks with clear community struc-
tures, which can be used to evaluate whether a graph layout
method preserves such structures in the layout. We use the
Stochastic Block model [24] implemented in the graph-tool [33]
to generate such graphs. The number of communities in the
generated graphs ranges from five to fifteen.
• Real graphs are networks publicly available on the Internet used
by prior studies [29, 35, 43]. They come from different fields such
as biology, social sciences and environmental structures, and have
different sizes.
We generated fifteen example graphs for each of the three types of
synthetic graphs and collected fifteen real graphs. To ensure that our
solver can efficiently find optimal solutions for our layout methods, we
did not test large graphs but used exemplars with 100 to 5000 nodes
and 128 to 19016 edges. Parameters. There are three parameters of
the augmented SGD solver, which influence the graph layout quality
Fig. 3. Comparing three solvers (BH, SGD, and our augmented SGD) for and speed, i.e., the maximum number of iterations, the step size η,
the FDP method to layout the btree9 graph. (a) The plots of U(X) versus and the decay rate of the Barnes-Hut algorithm λ . For η, we follow
the number of iterations showing the convergence of the three solvers. the suggestion of Zhang et al. [42] that gradually decreases from 1
Our augmented SGD solver performs similarly to SGD. (b,c,d) The layout to 0.01, while setting the maximum number of iterations to 200. For
results generated by three solvers, the one produced by our augmented SM and BSM, we found that usually 30 iterations are enough for
SGD solver preserves the tree structure well. convergence. To remove the influence of the initialization, we use the
same randomized initial layout for all methods on each graph.
5 E VALUATION Measures. Seven measures are used to evaluate the similarity of
We implemented our framework in C++ and warped it into an open- different implementations for the same graph layout method and
source library1 . To customize a graph layout method, users only need to to compare their performances. These measures were chosen to
configure and assemble different forces together in terms of the design evaluate the capability for different graph structure preservation or
guidelines from Section 3.3. Based on this library, we evaluate Taurus graph readability.
from four perspectives. First, we validate our augmented SGD solver • Normalized stress error (SE) [13] is used to measure the overall
by comparing it with BH and SGD on a set of graphs. Second, we check preservation of the graph-theoretical distances in the graph
to what extent the implementations of existing graph layout methods layouts. A small value indicates that graph-theoretical distances
under Taurus generate similar results as the original ones. Third, we between nodes are well maintained.
compare our proposed balanced stress model with the existing graph
layout methods. Last, we present a usage scenario to illustrate the • Neighborhood preservation (NP) checks whether the neighbor-
flexibility of Taurus for supporting the customization of graph layouts. hood around each node in the graph structure is also the neigh-
All the experiments were done on a Windows desktop computer with borhood in the layout. We use the neig hborhood preservation
an Intel Core i7-9700K processor with 32GB memory. measure introduced in [37], and define the neighborhood nodes
of a graph node as the nodes with a maximum of two edges from
5.1 Experimental Design
it. A larger value of NP is preferred.
Methods. To evaluate the capability of Taurus in expressing existing
graph layout techniques, we selected five well-known methods as • Crosslessness (CL) quantifies the number of non-crossing edge
baselines: force-directed placement (FDP) [10], LinLog [31], stress pairs in a graph layout [34]. A larger score indicates fewer edge
model (SM) [14], Maxent [13], and FM3 [18]. They are chosen by crossings.
considering two factors: First, they cover three major categories of • Minimum Angle (MA) measures the average deviation of the min-
existing graph layout methods (spring-electrical models: FDP, LinLog imum angle from the ideal angle for each node in a layout. [34] A
and FM3 ; stress models: SM; hybrid models: Maxent), as introduced in small value is preferred.
Section 2.1. Second, they have widely-used implementations. Four of
these methods have C++ implementations: FDP, FM3 and SM within • Runtime measures the average time for computing a layout. We
the OGDF [6] package, and Maxent within the graphviz [9] package. assess the average runtime of each graph layout method on one
For LinLog, we use the author’s Java implementation [31]. We re- graph by calculating the average over 5 runs.
implement these graph layout methods under Taurus, and compared
our results with the existing implementations. • Cluster Extraction (CE) [38] delineates the average distance of
Datasets. To evaluate our general framework, we generated three types nodes within the same cluster. A small value of this measure
of commonly-used graph datasets (i.e., grids, binary trees, and clustered indicates compact clusters, which is helpful for identifying graph
graphs) and 15 real graphs of different applications. communities and is thus preferred.
• Cluster Distance (CD) [38] measures the separation of different
• Grids are graphs with a regular tiling. The ideal layout results clusters in the layout. We calculate the minimum distance
will be a uniform grid consisting of squares of a uniform size. between nodes of two different clusters here. A large value
Grids can be used to assess whether a graph layout method is able indicates different clusters are well separated and is thus preferred.
to preserve the regular graph structures. We generate both 2D and
3D grids of different sizes. Since CE and CD both require the cluster information, we only apply
them to the clustered graphs.
• Binary trees have been widely used to evaluate graph layout To consistently compare different methods, we evaluate the relative
methods [13, 27]. Binary trees are often symmetric. Using difference of the measure M between the target implementation Mt and
them, we can evaluate whether a method is able to preserve the the reference implementation Mr :
symmetric structures of a graph.
Mt − Mr
1 https://github.com/Ideas-Laboratory/Taurus δM = , (14)
Mr
 Fig. 5. Relative differences of the measures SE(a), NP(b), CL(c), MA(d)
and runtime(e) for the layouts of all tested graphs and CD(f) and CE(g)
for the clustered graphs by our implementation vs. existing ones. A larger
value for all measures is better.
Fig. 4. (a) The boxplots of the runtime for different solvers, in which black
lines represent the median value. (b) the plots of runtime versus number
of nodes for each dataset. (c,d,e,f) The boxplots of four measures SE(c),
NP(d), CL(e) and MA(f) for the layouts generated by using ASGD vs. BH
and SGD. A large value is better in all boxplots.

where a value around zero indicates a specific implementation is

similar to the reference implementation. Among all measures, positive
differences of NP, CL and CD indicate a better performance of the
target implementation, while negative differences of SE, MA, CE and
runtime mean that the reference implementation performs better. To
consistently show larger values as being better, we take the final value
for SE, MA, CE and runtime as 1 − δ M.
5.2 Comparison between Different Solvers
Fig. 6. Comparison of layout results on the same data generated by
Since our ASGD is equivalent to SGD for stress model based methods, the same methods using the existing implementation (left) and our
we only compare its efficiency with SGD and BH for spring based implementation (right). (a,b) Results generated by FDP; (c,d) results
methods. Here, we choose the classic FDP method as an example generated by Maxent.
method for comparison in terms of runtime and layout quality. To
ensure a fair comparison, we use the same convergence condition for method under our framework as Mt with its original or existing
three solvers, while running each solver for a graph five times and implementation as Mr and compute δ M in terms of the above measures.
calculating the average measures. For all measures, we apply Eq. 14 to Quantitative Results. The boxplots in Fig. 5 summarize the relative
normalize the results by taking the implementation based on our ASGD differences of seven measures computed from all tested datasets. For
as Mt and the two others as the references. SE, the median values of the relative differences on all methods are
Due to space limits, we only show the summarized runtime and four either around zero (i.e., SM, Maxent, LinLog and FM3 with a difference
measures in Fig. 4, the complete scores of all measures can be found less than 3%), or larger than zero (FDP with a difference around 15%)
in the supplemental material, as well as runtimes and visualizations as shown in Fig. 5(a). Similarly, the median value of the relative
for additional ten large graphs. The boxplots in Fig. 4(a) provide a differences of NP on SM is around zero, the ones of Maxent, LinLog
statistical summary of the runtime of three methods on all tested graphs, and FM3 are larger than zero.
where our ASGD is faster than SGD by 40 percent and slightly slower These results indicate that our implementation enables FDP to better
than BH. To learn how fast these solvers are, we plot the relationship preserve stress and neighborhood, while showing similar performances
between runtime and the number of graph nodes in Fig. 4(b). We for the other methods. We have similar observations for the two
can see that all solvers perform similarly for graphs with a number of readability measures in Figs. 5(c,d). Our implementations result in
nodes being smaller than 1500, while ASGD and BH have significant less edge crossing (CL) and smaller minimum angle (MA) for Maxent,
advantages over SGD for large graphs. With increasing number of FDP LinLog, and FM3 , while maintaining the quality of SM.
graph nodes, the runtime of SGD increases quadratically, while the Furthermore, it largely reduces the runtime of SM, Maxent, FDP
runtime of ASGD and BH increases logarithmically. For the largest and LinLog by around 75%, 45%, 75% and 90%, respectively (see
graph with 5000 nodes, ASGD and BH are six times faster than SGD Fig. 5(e)). For FM3 , it leads to a slightly longer runtime (less than 5%)
(1.8s vs. 10.5s). for most graphs. We speculate that the reason or this is the fact that our
Figs. 4(c-f) provide a summary of four layout quality measures in ASGD solver is not inherently designed for solving multi-level graph
terms of relative differences. For SE and NP, ASGD demonstrates layout whereas the BH method is. Note that the SM implementation
significant benefits over SGD and BH, where the median of the relative in the OGDF library is solved by stress majorization rather than SGD,
increase is at least 8%, while it is similar for the other measures. In FDP in the OGDF library is the exact implementation, while the other
all, ASGD is significantly faster than SGD and similar to BH, while methods in the existing implementations are based on BH methods.
generating the best layout. Figs. 5(f,g) show the results of the measures CD and CE for clustered
5.3 Comparison between Different Implementations graphs. We can see that the relative differences of all methods
are positive or close to zero for CE and CD, indicating that our
To demonstrate the effectiveness of our quotient-based representation implementation has a higher capability in revealing cluster structures.
and ASGD solver, we compare the implementation of each layout
Fig. 7. Six measures over all datasets for six layout methods. A large value is better for all measures.

with regard to CE and CD on clustered graphs. This is reasonable, since

LinLog is designed for revealing clusters, while BSM is for distance
preservation.
While LinLog performs the worst in terms of the stress error, it is
the best for preserving neighborhoods. This is interesting, since we
often assume that LinLog is good in revealing clusters and FDP and
FM3 performs well in neighborhood preservation. After examining the
statistics of each type of graph in Fig. 8, we found that FM3 works
well for all graphs and LinLog performs poorly on grids but works
well for other graphs, whereas SM and our BSM perform best for such
graphs. For tree structured graphs, BSM performs better than SM but
worse than the others, because maintaining the overall distance is not
helpful here. In contrast, BSM performs similar or even better than
FDP and FM3 on clustered and real graphs, although it is designed for
preserving graph-theoretical distances. For each of the other measures,
Fig. 8. Bxplots summarizing the NP measure of each type of graphs for the distribution of the results does not show any significant dependency
six different layout methods. A larger value is better. on the data type.
In summary, BSM achieves a balanced capability in preserving
Qualitative Results. The above analysis shows that our implementa- overall distances (SE), neighborhoods (NP) and clusters (CE and CD),
tion produces similar results as the baseline for stress model and LinLog, especially for clustered and real graphs. In addition, it maintains similar
while largely improving the layout results for FDP and Maxent. Fig. 6 graph readability values as the other methods and therefore matches
compares the results produced by different implementations of FDP our design goal.
and Maxent on four typical graphs. Qualitative Results. Fig. 9 shows the visual results of four typical
Our FDP implementation is able to effectively maintain the grid example graphs. The dataset in the first row of Fig. 9 is a graph with
structure for the graph shown in Fig. 6(a), whereas the grid is highly a 3D-like grid, where only BSM and SM preserve the grid structure,
deformed with folds by the baseline implementation. We have similar while the other methods heavily deform it.
observations for the tree structure shown in Fig. 6(b), our result on the BSM is able to clearly visualize the major branches and the
right has fewer edge crossings and reveals a clearer hierarchy of clusters symmetry of tree graph shown in the second row of Fig. 9. Other
than the original one on the left. These results are consistent with the methods (except SM) do not preserve the symmetry but show more
ones shown in Fig. 3, demonstrating the efficiency of our solver. minor branches. For this graph, Maxent, FDP and LinLog have a
Figs. 6(c,d) show results generated by two implementations of higher score in neighborhood preservation, whereas they do not show
Maxent. Our result is significantly better in terms of cluster preservation major branches. Hence, we speculate that BSM might perform even
and reveals all different communities (Fig. 6(c)), while the original better in visualizing large tree structures than these methods. Although
mixes clusters. This also holds for the collaboration network (Fig. 6(d)). LinLog allows to inherently reveal cluster structures, its results are
Compared to the baseline, our result distributes all local clusters in 2D too tight to explore any details. FDP and Maxent alleviate this issue,
plane evenly with fewer edge crossing, while maintaining the overall but might create overlaps between clusters. In contrast, BSM better
structure. Yet, the baseline implementation better reveals the multi-ring balances intra- and inter-cluster separation well, where the maroon
structures, but has strong node overlaps and edge crossings. and yellow green clusters are at the center and are surrounded by the
Overall, the results of all methods implemented in Taurus are similar other clusters. This makes sense, since these two clusters have the most
to or even better than the original implementations, with a largely inter-cluster connections with the others in the ground truth data. In
reduced runtime. In particular, our versions of FDP and Maxent largely contrast, FM3 does not reflect this ground truth. For the real graphs, the
improve layout quality. results of BSM are similar to or even slightly better than SM in terms
of preserving radial graph structures and local clusters, while both of
5.4 Comparison between Different Methods them perform better than the others.
Here, we compare our proposed layout method BSM with five Overall, the visual results by BSM align well with the observations
existing methods (SM, FDP, Maxent, LinLog and FM3 ). To show in Section 5.4: BSM is able to visualize the structures of different types
the effectiveness of this model, we implemented all these methods by of graphs well.
Taurus with the same solver and use the same initial layout for each
tested graph. 5.5 Usage Scenario
Quantitative Results. Fig. 7 summarizes the values of six measures Lastly, we present a usage scenario to showcase how Taurus can
overall tested graphs using boxplots. For the complete results presented facilitate graph visualizations in a unified manner. Suppose Bob, a
in table, please refer to the supplemental material In terms of the data scientist who often needs to explore networks in his daily work
stress error, BSM is slightly worse than SM but outperforms the and has a basic understanding of different graph layout techniques.
other methods, while LinLog is the worst. In contrast, BSM is better He wants to visualize a network that he wants to explore the potential
than SM and Maxent in neighborhood preservation, close to FDP clusters. With Taurus, he can easily try different graph layout methods
and FM3 , worse than LinLog. Regarding readability measures, BSM within the same framework.
performs similarly to the other methods but is worse than LinLog for First, he uses Taurus to quickly implement BSM by simply
CL. Similarly, BSM is worse than LinLog and similar as FDP and FM3 configuring the four parameters {Ω, ω, α, β } for attractive forces with
Fig. 9. BSM results of four graphs (top down: grid_1000, btree_513, cluster_4463, and US_powergrid) with different structures in comparison to
baseline methods (SM, FDP, LinLog, Maxent and FM3 ) implemented by Taurus.

6 C ONCLUSION AND F UTURE W ORK

In this paper, we propose a general framework, Taurus, to unify popular
graph layout methods. It consists of two major components: a unified
quotient-based force representation to model repulsive and attractive
forces of different graph layout techniques, and a universal augmented
stochastic gradient descent (ASGD) solver to find the optimal graph
layout results. We systematically analyze our general framework and
provide guidelines for designing effective graph layout methods. We
also release a graph layout library based on Taurus that facilitates
convenient implementation of graph visualizations in a unified manner.
In the future, we would like to extend Taurus along the following
Fig. 10. Utilizing Taurus to obtain good graph layout results. By entering directions. First, apart from the current open-source package in
the corresponding parameters, the user quickly gets layout results for (a) C++, we plan to extend Taurus to other programming languages
BSM, (b) LinLog and (c) a customized layout method. like Python and JavaScript. Second, Taurus provides a clear design
space for graph layout techniques, but it might require users to try
{V 2 , 1, 1, 1} and for repulsive forces with {V 2 , −1, −1, −1}. The multiple different parameters for a specific graph analysis task (see
layout result (Fig. 10 (a)) shows that the network seems to have a Section 5.5). Therefore, we like to explore automated parameter
clustering structure. To display clusters more clearly, Bob updates tuning methods for automatically fining proper parameters to generate
the parameters to {E, 1, 0, 0} and {V 2 , −1, −1, 0} for attractive and desired visualizations. Third, it will be interesting to further extend
repulsive forces, i.e. applying the LinLog method. He obtains a more our framework to unify graph layout techniques with special model
compact clustering result as shown in Fig. 10 (b). However, the clusters designs (e.g., tsNET [29] and DRGragh [43]). Last, we would like to
in this graph layout are too tight to examine the relationship between conduct a large user study to compare the different layout methods in
nodes within the same cluster. To achieve the optimal graph layout, terms of layout principles and then use the findings to further improve
Bob further adjusts α and β of the attractive force to 2.5 and -1.5, and our framework.
finds that the result in Fig. 10 (c) shows the detail of each cluster more
clearly, making it easier for Bob to know which clusters contain more ACKNOWLEDGMENTS
nodes. With Taurus, Bob can easily customize graph layout methods to This work was supported by the grants of the National Key Research
explore different graph structures, which facilitates him to perform a and Development Program of China (2018AAA0101100), and NSFC
deeper analysis. (62132017, 62141217). Oliver Deussen was funded by the German
Research Foundation (DFG) - Project-ID 2,51654672 - TRR 161.
R EFERENCES [24] B. Karrer and M. E. Newman. Stochastic blockmodels and community
[1] R. Ahmed, F. De Luca, S. Devkota, S. Kobourov, and M. Li. Multicriteria structure in networks. Physical Review E, 83(1):016107, 2011. doi: 10.
scalable graph drawing via stochastic gradient descent, (SGD)2 . IEEE 1103/PhysRevE.83.016107
Transactions on Visualization and Computer Graphics, 28(6):2388–2399, [25] M. Kaufmann and D. Wagner. Drawing graphs: methods and models.
2022. doi: 10.1109/TVCG.2022.3155564 Springer, 2003.
[2] D. Auber, D. Archambault, R. Bourqui, M. Delest, J. Dubois, A. Lambert, [26] A.-M. Kermarrec and A. Moin. FlexGD: A flexible force-directed model
P. Mary, M. Mathiaut, G. Melançon, B. Pinaud, B. Renoust, and J. Vallet. for graph drawing. In IEEE Pacific Visualization Symposium, pp. 217–224.
TULIP 5. In Encyclopedia of Social Network Analysis and Mining, pp. IEEE, 2013. doi: 10.1109/PacificVis.2013.6596148
1–28. Springer, 2017. doi: 10.1007/978-1-4614-7163-9_315-1 [27] M. Khoury, Y. Hu, S. Krishnan, and C. Scheidegger. Drawing large graphs
[3] J. Barnes and P. Hut. A hierarchical o (n log n) force-calculation algorithm. by low-rank stress majorization. Computer Graphics Forum, 31(3pt1):975–
Nature, 324(6096):446–449, 1986. doi: 10.1038/324446a0 984, 2012. doi: 10.1111/j.1467-8659.2012.03090.x
[4] M. Bostock, V. Ogievetsky, and J. Heer. D3 data-driven documents. IEEE [28] S. G. Kobourov. Force-directed drawing algorithms. 2004.
Transactions on Visualization and Computer Graphics, 17(12):2301–2309, [29] J. F. Kruiger, P. E. Rauber, R. M. Martins, A. Kerren, S. Kobourov,
2011. doi: 10.1109/tvcg.2011.185 and A. C. Telea. Graph layouts by t-SNE. Computer Graphics Forum,
[5] U. Brandes and C. Pich. Eigensolver methods for progressive 36(3):283–294, 2017. doi: 10.1111/cgf.13187
multidimensional scaling of large data. In International Symposium on [30] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature,
Graph Drawing, pp. 42–53. Springer, 2006. doi: 10.1007/978-3-540 521(7553):436–444, 2015. doi: 10.1038/nature14539
-70904-6_6 [31] A. Noack. An energy model for visual graph clustering. In International
[6] M. Chimani, C. Gutwenger, M. Jünger, G. W. Klau, K. Klein, and symposium on graph drawing, pp. 425–436. Springer, 2003. doi: 10.
P. Mutzel. The open graph drawing framework OGDF. Handbook of 1007/978-3-540-24595-7_40
Graph Drawing and Visualization, 2011:543–569, 2013. [32] M. Ortmann, M. Klimenta, and U. Brandes. A sparse stress model. In
[7] T. Dwyer. Scalable, versatile and simple constrained graph layout. International Symposium on Graph Drawing and Network Visualization,
Computer Graphics Forum, 28(3):991–998, 2009. doi: 10.1111/j.1467 pp. 18–32. Springer, 2016. doi: 10.1007/978-3-319-50106-2_2
-8659.2009.01449.x [33] T. P. Peixoto. The graph-tool python library. figshare, 2014. doi: 10.
[8] P. Eades. A heuristic for graph drawing. Congressus Numerantium, 6084/m9.figshare.1164194
42:149–160, 1984. [34] H. C. Purchase. Metrics for graph drawing aesthetics. Journal of Visual
[9] J. Ellson, E. Gansner, L. Koutsofios, S. C. North, and G. Woodhull. Languages & Computing, 13(5):501–516, 2002. doi: 10.1006/jvlc.2002.
Graphviz—open source graph drawing tools. In International Symposium 0232
on Graph Drawing, pp. 483–484. Springer, Feb. 2001. doi: 10.1007/3-540 [35] R. A. Rossi and N. K. Ahmed. The network data repository with interactive
-45848-4_57 graph analytics and visualization. Twenty-Ninth AAAI Conference on
[10] T. M. Fruchterman and E. M. Reingold. Graph drawing by force-directed Artificial Intelligence, 29(1):4292–4293, 2015. doi: 10.1609/aaai.v29i1.
placement. Software: Practice and Experience, 21(11):1129–1164, 1991. 9277
doi: 10.1007/978-3-658-21742-6_49 [36] R. Tamassia. Handbook of Graph Drawing and Visualization. CRC press,
[11] P. Gajer, M. T. Goodrich, and S. G. Kobourov. A multi-dimensional 2013.
approach to force-directed layouts of large graphs. Computational [37] L. Van der Maaten and G. Hinton. Visualizing data using t-SNE. Journal
Geometry, 29(1):3–18, 2004. doi: 10.1016/j.comgeo.2004.03.014 of Machine Learning Research, 9(86):2579–2605, 2008.
[12] P. Gajer and S. G. Kobourov. Grip: Graph drawing with intelligent [38] F. Van Ham and B. Rogowitz. Perceptual organization in user-generated
placement. In International Symposium on Graph Drawing, pp. 222–228. graph layouts. IEEE Transactions on Visualization and Computer
Springer, May 2000. doi: 10.1142/9789812796608_0011 Graphics, 14(6):1333–1339, 2008. doi: 10.1109/TVCG.2008.155
[13] E. R. Gansner, Y. Hu, and S. North. A maxent-stress model for graph [39] C. Walshaw. A multilevel algorithm for force-directed graph drawing.
layout. IEEE Transactions on Visualization and Computer Graphics, In International Symposium on Graph Drawing, pp. 171–182. Springer,
19(6):927–940, 2012. doi: 10.1109/TVCG.2012.299 2000. doi: 10.1142/9789812773296_0012
[14] E. R. Gansner, Y. Koren, and S. North. Graph drawing by stress [40] Y. Wang, Y. Wang, Y. Sun, L. Zhu, K. Lu, C.-W. Fu, M. Sedlmair,
majorization. In International Symposium on Graph Drawing, pp. 239– O. Deussen, and B. Chen. Revisiting stress majorization as a unified
250. Springer, 2004. doi: 10.1007/978-3-540-31843-9-25 framework for interactive constrained graph visualization. IEEE
[15] E. R. Gansner and S. C. North. An open graph visualization system and its Transactions on Visualization and Computer Graphics, 24(1):489–499,
applications to software engineering. Software: Practice and Experience, 2017. doi: 10.1109/TVCG.2017.2745919
30(11):1203–1233, 2000. [41] H. D. Young, R. A. Freedman, T. Sandin, and A. L. Ford. University
[16] A. M. Gouvêa, T. S. da Silva, E. E. Macau, and M. G. Quiles. Force- Physics, vol. 9. Addison-Wesley Reading, MA, 1996.
directed algorithms as a tool to support community detection. The [42] J. X. Zheng, S. Pawar, and D. F. Goodman. Graph drawing by stochastic
European Physical Journal Special Topics, 230(14):2745–2763, 2021. gradient descent. IEEE Transactions on Visualization and Computer
doi: 10.1140/epjs/s11734-021-00167-0 Graphics, 25(9):2738–2748, 2018. doi: 10.1109/TVCG.2018.2859997
[17] R. Gove. A random sampling o (n) force-calculation algorithm for graph [43] M. Zhu, W. Chen, Y. Hu, Y. Hou, L. Liu, and K. Zhang. Drgraph: An
layouts. Computer Graphics Forum, 38(3):739–751, 2019. doi: 10.1111/ efficient graph layout algorithm for large-scale graphs by dimensionality
cgf.13724 reduction. IEEE Transactions on Visualization and Computer Graphics,
[18] S. Hachul and M. Jünger. Drawing large graphs with a potential-field-based 27(2):1666–1676, 2020. doi: 10.1109/TVCG.2020.3030447
multilevel algorithm. In International Symposium on Graph Drawing, pp.
285–295. Springer, 2004. doi: 10.1007/978-3-540-31843-9_29
[19] S. Hachul and M. Jünger. Large-graph layout with the fast multipole
multilevel method. Spring, V (December), pp. 1–27, 2005.
[20] Y. Hu. Efficient high-quality force-directed graph drawing. Mathematica
Journal, 10(1):37–71, 2005.
[21] Y. Hu and Y. Koren. Extending the spring-electrical model to overcome
warping effects. In IEEE Pacific Visualization Symposium, pp. 129–136.
IEEE, May 2009. doi: 10.1109/PACIFICVIS.2009.4906847
[22] M. Jacomy, T. Venturini, S. Heymann, and M. Bastian. Forceatlas2,
a continuous graph layout algorithm for handy network visualization
designed for the gephi software. PLOS ONE, 9(6):e98679, 2014. doi: 10.
1371/journal.pone.0098679
[23] T. Kamada, S. Kawai, et al. An algorithm for drawing general undirected
graphs. Information Processing Letters, 31(1):7–15, 1989. doi: 10.1142/
9789814434478_0005