EDITOR: Mike Potel, potel@wildcrest.

com

DEPARTMENT: APPLICATIONS

Evaluating Representation Learning and

Graph Layout Methods for Visualization
Edith Heiter , Bo Kang , Tijl De Bie , and Jefrey Lijffijt , Department of Electronics and Information
Systems, IDLab, Ghent University, 9000, Ghent, Belgium

Graphs and other structured data have come to the forefront in machine learning over
the past few years due to the efficacy of novel representation learning methods
boosting the prediction performance in various tasks. Representation learning
methods embed the nodes in a low-dimensional real-valued space, enabling the
application of traditional machine learning methods on graphs. These representations
have been widely premised to be also suited for graph visualization. However, no
benchmarks or encompassing studies on this topic exist. We present an empirical
study comparing several state-of-the-art representation learning methods with two
recent graph layout algorithms, using readability and distance-based measures as
well as the link prediction performance. Generally, no method consistently
outperformed the others across quality measures. The graph layout methods provided
qualitatively superior layouts when compared to representation learning methods.
Embedding graphs in a higher dimensional space and applying t-distributed
stochastic neighbor embedding for visualization improved the preservation of local
neighborhoods, albeit at substantially higher computational cost.

V
isualization of data is useful both for under- finding planar embeddings (no edge crossings), and
standing data and cross-checking models later research also considered embeddings of large
trained on that data. Consequently, it may graphs, leading to the development of the still popu-
lead to new insights, better models, the detection of lar force-directed approaches. In contrast, graph
outliers, etc. Visualization is frequently used in data representation learning methods (sometimes also
science. However, it is not straightforward to visualize referred to as network embedding methods) have not
a graph. The vertices and edges typically do not have been developed specifically for visualization, but
a real-valued representation, hence, the primary prob- rather they aim to embed graphs into a low-dimen-
lem in graph visualization is to find a good representa- sional space (depending on the method, between 8
tion of the vertices in two-dimensional space. and 128 dimensions in general), to enable subsequent
Traditionally, this area was known as graph draw- use of common machine learning techniques.
ing and methods that produce a representation of There exist many graph-layout algorithms, each
the nodes of a graph on a two-dimensional real-val- with a different objective or optimization procedure.
ued space were referred to as graph-layout algo- Someone who wants to inspect or analyze a graph
rithms. Early research considered problems such as visually is then already faced with a difficult choice of
which method would be most appropriate for their
goal. This difficulty is amplified by the introduction of
dozens of representation learning methods for graphs,
which have been premised to also produce good visu-
This work is licensed under a Creative Commons Attribution alizations (see, e.g., the works of Perozzi et al.1 and
4.0 License. For more information, see https://creativecom- Tsitsulin et al.2).
mons.org/licenses/by/4.0/
Digital Object Identifier 10.1109/MCG.2022.3160104
It is not clear which method to use, or even which
Date of current version 8 June 2022. aspects vary by choosing one method instead of

May/June 2022 Published by the IEEE Computer Society IEEE Computer Graphics and Applications 19
APPLICATIONS

FIGURE 1. Visualizations of the netscience collaboration graph using t-SNE and different embedding methods: (a) AROPE,
(b) CNE, (c) DeepWalk, (d) GAE. Color corresponds to edge length: yellow (long) to dark blue (short).

another. For example, in Figure 1, we show visualiza- Quality Measures

tions from different representation learning methods. There exist a variety of scores to measure the quality
It is not obvious which visualization is better than of graph layouts. The graph representation learning
another or how they differ. community typically evaluates graph embeddings on
Representation learning methods are often evalu- tasks, such as clustering, link prediction, graph recon-
ated by coloring the vertices according to a class label, struction, or node classification. These are inspired by
and a good visualization is then one where nodes from and suitable to evaluate embeddings as models of the
the same class are (visually) grouped together. But data, either considering the performance of an embed-
such evaluations have limited scope, because whether ding toward reconstruction or prediction. In contrast,
an algorithm does well depends strongly on what type recent methods for graph visualization (see the works
of structure from the graph aligns with the node labels. of Kruiger et al.,3 Zhu et al.,4 and Kwon5) are mostly
New graph layout methods are evaluated on more evaluated by means of readability, shape-based, or dis-
immediate quality measures, but are not compared to tance-based measures. Such measures inform more
representation learning methods (see, e.g., the works directly about the quality of a graph layout, as they are
of Kruiger et al.3 and Zhu et al.4). It is an open question easy to interpret, but they may not capture to what
how to generally evaluate graph layouts and whether extent we can do inference and prediction using a layout.
some currently used methods outperform others. As the measures may be complementary to each
Contributions: In this article, we perform an empiri- other, we evaluate graph layouts on measures from
cal analysis of graph visualizations by representation each of these groups. We quantify the readability
learning and graph layout algorithms. We investigate using crosslessness, minimum angle, edge length vari-
the differences in their visualizations visually and quan- ation, and Gabriel-shape. Neighborhood preservation
titatively using a diverse set of quality measures. Repre- and stress are distance-based measures. We also
sentation learning methods are often used to compute quantify the suitability of the layouts for link predic-
node embeddings in more than two dimensions. We, tion. We repeat the embedding process four times to
thus, evaluate native two-dimensional and also higher reduce random effects and use the graph layout aes-
dimensional node representations that we embed into thetic metrics (GLAM) toolbox5 to compute the four
two dimensions using the well-known t-distributed sto- readability measures.
chastic neighbor embedding (t-SNE) method. We base We next introduce these measures and use the fol-
our analysis on publicly available methods and provide lowing notation. We denote an undirected, unweighted,
the full source code, which can be easily extended to and connected graph G ¼ ðV; EÞ by the set of nodes V
include other methods or datasets. Lastly, we give rec- with jV j ¼ n and the set of edges E ¼ fði; jÞg  V  V .
ommendations on how to apply t-SNE to retain impor- We refer to the graph-theoretical shortest-path dis-
tant graph structure and how different methods could tance between nodes i and j by dij . With Nði; hÞ ¼ fj 2
be improved to yield better visualizations. V jdij  h; i 6¼ jg ; we denote the set of nodes in the
h-order neighborhood of node i. With the following
readability measures, we evaluate two-dimensional
SETUP OF THE STUDY graph layouts Y 2 Rn2 , where the embedding of a
In this section, we introduce our choice of quality node i is denoted as yi 2 R2 and the edges are drawn
measures, graphs, and evaluated methods. using straight lines.

20 IEEE Computer Graphics and Applications May/June 2022

 APPLICATIONS

Crosslessness6 measures the proportion of edge

1 X N
jNG ði; hÞ \ NY ðyi ; ki Þj
crossings that are avoided in a graph layout with
jV j i¼1 jNG ði; hÞ [ NY ðyi ; ki Þj
respect to all possible pairwise edge crossings. We
adopt the definition by Kwon et al.5 where NG ði; hÞ is the graph neighborhood of node i

c with distance at most h and NY ðyi ; ki Þ are the ki ¼
1 cmax cmax > 0
crosslessness ¼ jNG ði; hÞj nearest neighbors based on Euclidean dis-
1 Otherwise
tances in the embedding. We evaluate the neighbor-
where c is the number of crossings and cmax the upper hood preservation for h 2 f1; 2g.
bound on the number of crossings. cmax is defined as Stress measures how accurately a graph layout
jEjðjEj1Þ P represents the actual graph-theoretic node distances.
2  12 i2V jNði; 1ÞjðjNði; 1Þj  1Þ; where jNðiÞj is
the degree of node i. We define the stress similar to Zhu et al.4 as
Minimum angle6 measures how much the smallest 2 X akyi  yj k  dij 2
angle between incident edges at each node deviates jV jðjV j  1Þ i < j2V dij
from the optimal angle that could be achieved when
all edges are evenly spaced where a scales the embedding to minimize the stress
(see the supplementary material, which is available
1 X juðiÞ  umin ðiÞj
minimum angle ¼ 1  at http://doi.ieeecomputersociety.org/10.1109/MCG.
jV j i2V uðiÞ
2022.3160104.).
where uðiÞ ¼ jNðiÞj
360
is the optimal angle at node i and Link prediction is the task of predicting links miss-
umin ðiÞ is the minimum angle between any two edges ing from a graph. We use the EvalNE toolbox9 to evalu-
incident to node i. ate the graph layouts for this task. A random
Edge length variation7 measures the variability of connected subset of 80% of the edges is used to com-
the edge lengths relative to their mean. It is defined as pute the layout and the remaining 20% together with
the coefficient of variance the same number of randomly sampled nonexisting
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi edges are used for testing. While some methods have
P 2
ls e2E ðle  lm Þ built-in edge embeddings [arbitrary-order proximity
lcv ¼ ¼ :
lm jEj  l2m preserved network embedding (AROPE), CNE, GAE],
others only result in node embeddings. In the latter
To be in line with the other readability measures where case, we use EvalNE to transform the node embed-
higher values are better, we define dings into edge embeddings using their element-wise
lcv average, Hadamard product, L1 , or L2 distance. We
Edge length uniformity ¼ 1  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi then predict the test edge probabilities with a logistic
jEj  1
classifier fitted on the train edge embeddings and
where the coefficient of variance is normalized by its report the AUC-ROC.
upper bound.
Gabriel-shape similarity8 is a shape-based quality Datasets
measure for layouts of large graphs. It is based on the We evaluate visualizations of the largest connected
intuition that a representation of the global graph components of six undirected and unweighted graphs
structure can be accurate despite a high number of listed in Table 1. The karate club graph is a commonly
edge crossings. It is defined as the mean Jaccard simi- used example and models the social interactions of
larity the sports club members. can96 is an artificial mesh
1 X jNG ði; 1Þ \ NS ði; 1Þj structure, netscience a coauthorship graph from net-
jV j i2V jNG ði; 1Þ [ NS ði; 1Þj work scientists, and powergrid represents the electric-
ity grid of the western United States. The Facebook
between the first-order neighborhood in the original friendship graph is commonly evaluated on the link
graph NG and the neighborhood in the Gabriel-shape prediction task. The largest graph, a Twitter feed graph,
graph NS . The Gabriel graph over the set of nodes V has been used in the experiments by Zellmann et al.10
has an edge connecting two nodes i and j if a disk and provided to us by the authors.
with yi and yj on its boundary and diameter kyi  yj k
does not contain any other node. Evaluated Methods
Neighborhood preservation3 is defined as the Jac- In this study, we compare graph visualizations from
card similarity seven methods, listed in Table 2 with parameter

May/June 2022 IEEE Computer Graphics and Applications 21

APPLICATIONS

TABLE 1. Characteristics of graphs. TABLE 2. Overview of evaluated methods.

Name jV j jEj Type Source

11
karate 34 78 Social network
club
11
can96 96 429 Artificial network
11
netscience 379 914 Coauthorship
graph
12
facebook 4039 88 234 Social network
11
powergrid 4941 6594 Infrastructure
graph
10
twitter 44 631 73 133 Social network

settings and full URLs listed in Table 11 of the supple-

mentary material, available online. We selected four
graph representation learning methods: AROPE, CNE,
DeepWalk, and GAE, which are state-of-the-art meth-
ods each using a distinct approach.17 Those methods
are mostly used to compute low-dimensional node
embeddings in more than two dimensions and the
two-dimensional graph layouts for visualization are FR16 is a force-directed approach with attraction
then obtained by t-SNE.18 We denote these node forces between connected nodes and repulsion
embeddings by, e.g., AROPE128, indicating the between all nodes. Node positions are updated itera-
dimensionality before applying t-SNE. We also evalu- tively with a temperature cooling scheme that reduces
ate native two-dimensional node representations for the displacement of nodes over time. We compare the
all graph representation learning methods denoted by, OðjV j2 Þ implementation of FR by Hagberg et al.19 (ver-
e.g., AROPE2. sion 2.2) with a GPU-optimized implementation by
The representatives for graph layout methods are Zellmann et al.,10 denoted as FR-RTX.
the Fruchterman–Reingold (FR) algorithm and DRGraph. DRGraph4 preserves the first-order neighbors from
We chose FR due to its popularity and evaluate its naive the graph in the low-dimensional representation by
implementation as well as a recent more scalable grid- minimizing the difference between node similarities
based optimization. DRGraph is a recently published and layout proximities. It is based on the idea of
method that is faster than existing layout techniques tsNET3 but improved the computational and memory
but produces visualizations with comparable readability complexity by introducing a sparse distance matrix,
and neighborhood preservation. negative sampling, and a multilevel optimization
AROPE13 preserves higher order proximities by fac- process.
torizing a linear combination of powers of the adja-
cency matrix. It is efficient through the use of eigen-
EMPIRICAL RESULTS
decomposition reweighting.
We first compare the graph layouts qualitatively based
Conditional network embedding (CNE)14 is a proba-
on the visualizations in Table 3 and then evaluate the
bilistic network embedding method based on maxi-
quantitative measures. The code and the embeddings
mum likelihood estimation that can factor out prior
are available at https://github.com/aida-ugent/graph-
knowledge about the graph structure using a Bayesian
vis-eval. The supplementary material, available online,
approach.
contains node-link diagrams (see Table 12) and scores
DeepWalk1 captures the neighborhood structure of
for all layouts (Tables 1– 10).
nodes using random walks. The node embeddings are
updated using the Skip-gram model, maximizing the
co-occurrence probability of nodes in the same Visual Inspection of the Embeddings
neighborhood. AROPE: The first embedding coordinate of graph lay-
Graph autoencoders (GAEs)15 embed the graph outs by AROPE2 is proportional to the eigenvector
into a latent space using a graph convolutional net- centrality of the nodes, making the embedding unlike
work encoder and an inner product decoder. the others. Central nodes may be identified on one

22 IEEE Computer Graphics and Applications May/June 2022

 APPLICATIONS

TABLE 3. Visualizations based on node embeddings showing only the edges.

May/June 2022 IEEE Computer Graphics and Applications 23

APPLICATIONS

FIGURE 2. Placement of low-degree nodes (dark) around a high-degree hub node (yellow, marked with red arrow) for the Twitter
graph. While nodes are clustered according to their degree in AROPE128 and GAE16, and to some extend in CNE16, leaf nodes
are close to their parent in DW128. (a) AROPE128. (b) CNE16. (c) DW128. (d) GAE16.

side and leaf nodes on the other side of the embed- more compact and nodes are pushed away from the
ding for karate club, can96, and powergrid. The nodes center. For FR-RTX , nodes are distributed more evenly
for the other three graphs are mostly aligned along around a shared connection, which hinders the forma-
two axes, thus hiding most of their connections. tion of visible clusters for powergrid but improves the
Embeddings by AROPE128 reveal more details of the readability of Twitter. We assume this is the effect of
graph structure. In the Twitter graph layout in Figure 2(a), approximating the repulsive forces in FR-RTX, causing
we observe a clustering according to node-degree. only the closest nodes to repel each other.
CNE: The native two-dimensional embeddings of GAE: The embeddings from GAE2 have a distinct
CNE2 exhibit a proximity-based arrangement of the circular shape due to the inner product decoder. High-
nodes where hub nodes are placed in the center of degree node embeddings have large coordinates and
their connections. The embeddings by CNE16 are simi- low-degree nodes are placed near the origin. In this
lar to AROPE128, DW128, and GAE16. For the Twitter graph layout, we can easily identify the most central
graph, we find the t-SNE embedding is more readable hub nodes, (e.g., for Facebook or Twitter). The embed-
as it shows cluster structure. We see in Figure 2(b) dings by GAE16 and AROPE128 have a similar local
that CNE16 also clusters nodes by degree. structure, as shown in Figure 2(d). In both embeddings,
DeepWalk: The embeddings by DW2 conceal the nodes from the same cluster are arranged by degree.
underlying shape of the graph as nodes are mostly
arranged on a curved line. DW128 produces readable
embeddings with a clear cluster structure. In Figure 2(c), Readability Measures
we note that DW128 embeds low-degree nodes close to In Figure 3, we show the scores for crosslessness, edge-
their connections resulting in a star shape around the length uniformity, minimum angle, and Gabriel shape,
hub node. averaged over four runs. Averaging over all graph data-
DRGraph: The graph layouts by DRGRAPH for nets- sets, the graph layout methods outperform the represen-
cience and powergrid are appealing and very similar to tation learning methods (see the supplementary
the layouts by DW128. The different communities of the material, available online, Table 10). DW128 achieves high
Facebook graph are well visible, but in the node-link dia- scores for the layouts of larger graphs. The high mini-
gram (see the supplementary material, available online, mum-angle scores of the layouts by FR-RTX and DW128
Table 12), we notice some long edges that have a leaf stem from the star-shaped arrangement of hub-nodes
node on one end. Long edges also dominate the visuali- and their connections [see Figure 2(c)]. The layouts by
zation of the Twitter graph and make it difficult to AROPE2, DW2, and GAE2 generally have small angles
observe any structure in the center of the layout. We do between incident edges as they optimize the dot-prod-
not know whether the parameter settings or the fact uct similarity of connected nodes. Furthermore, we
that DRGRAPH only preserves first-order graph distan- observe that AROPE2, DW2, and GAE2 score poorly for
ces cause the “hairball” structure of this visualization. the Gabriel shape measure. These methods optimize the
FR: Both implementations result in graph layouts node embeddings for dot-product similarity whereas the
with similar global structure but different local node Gabriel shape retrieves neighbors based on Euclidean
arrangement. For FR, we observe that clusters are distance.

24 IEEE Computer Graphics and Applications May/June 2022

 APPLICATIONS

FIGURE 3. Average readability scores for each dataset, ordered by number of nodes. Crosslessness and edge length uniformity
are scaled to better show the relative differences between methods. (a) Crosslessness. (b) Edge length uniformity. (c) Minimum
angle. (d) Gabriel shape similarity.

Distance-Based Measures than two datasets. Averaging over all datasets, the
We show the scores for second-order neighborhood graph layouts based on the FR algorithm result in the
preservation and stress in Figure 4(a) and (b), and the highest link predict scores. While it is difficult to iden-
scores of the first-order neighborhood preservation tify much structure in the hairball-shaped Twitter
measure in the supplementary material, available graph layouts by FR and DRGRAPH, they score highly
online; they are highly similar to the Gabriel shape on the link-prediction task.
measure. DW128 and DRGRAPH preserve the second-
order neighborhood best. Considering all methods,
the neighborhoods of the Facebook graph are better Runtime
preserved than the neighborhoods of powergrid or The average runtime to embed the whole graphs using
Twitter. We presume that the community structure of an Intel Xeon CPU E5-2620 v4 @ 2.10 GHz with one
the graph aligns well with the second-order neighbor- GeForce GTX 1080Ti is depicted in Figure 4(d). We ran
hoods. FR places these communities far apart and, the experiments for GAE on Twitter on a different
thus, achieves the highest score. The differences in machine with 256 GB RAM resulting in mean runtimes
the stress measure are subtle. Averaged over all data- of 5020 s for GAE2 and 4635 s for GAE16.
sets, FR and FR-RTX result in layouts with the lowest We notice that AROPE2 is the fastest method and
stress. The t-SNE-based embeddings from AROPE128, that the runtime increase for AROPE128 is mainly
DW128, and GAE16 are more distance faithful than caused by t-SNE, which took about 22, 43, and 660 s
their native two-dimensional node embeddings but to reduce the dimensionality from 128 to 2 for
the opposite holds for CNE2 and CNE16. Facebook, powergrid, and Twitter, respectively. FR-
RTX has a slightly larger runtime than the other meth-
ods on the smaller graphs possibly due to a small
Link-Prediction startup cost for a user-interface. Notably, on the
Results are presented in Figure 4(c). No single method Twitter dataset, FR-RTX runs in less than a minute,
achieves the highest link prediction score on more while FR takes almost 24 h. CNE16 has lower runtime

May/June 2022 IEEE Computer Graphics and Applications 25

APPLICATIONS

FIGURE 4. Average neighborhood preservation, stress, AUC-ROC, and runtime over four experiment repetitions. We scale the
AUC values to better show the relative differences. (a) Second-order neighborhood preservation. (b) Stress. (c) Link prediction
AUCROC. (d) Runtime in seconds on a logarithmic axis.

than CNE2 on the larger graphs as the optimization of worse on these measures. Although other distances
the 16-dimensional node representations stabilizes than Euclidean may be interpretable, it is not obvious
earlier than the two-dimensional representations. to what extent that may be the case. In addition to
the quality measures, the standard version of t-SNE
also defines high- and low-dimensional neighborhoods
DISCUSSION based on Euclidean distance. To retain the graph
In this study, we have shown that visualizations by neighborhoods, we would have to adjust the similarity
graph layout methods scored higher on the chosen definition of t-SNE.
quality measures than the native two-dimensional
node embeddings by representation learning meth-
ods. The combination of DeepWalk with t-SNE resu- Recommendations for Practical Use
lted in visualizations with the best local neighborhood It is important to note that the choice of method is not
preservation and highest scores in Gabriel shape but universal and depends on the task for which the visu-
does not scale to larger graphs. We believe that there alization is used. The designer of the visualization
is great potential in comparing a wider range of scal- should ask the question, which quality measure is
able graph layout and representation learning meth- most important to judge the effectiveness of the visu-
ods on real-world graphs with millions of nodes. alization, for that task. For example, out-of-sample
The standard definitions of Gabriel shape, neigh- quality measures like link prediction performance may
borhood preservation, and stress all assume that the not be important in the context when a static graph is
graph-theoretical distances are reflected by the to be analyzed that does not have any missing edges.
Euclidean distances in the low-dimensional embed- In other contexts, generalizability of the proximity of
ding. Methods, such as AROPE, DeepWalk, and GAE, nodes, for which link prediction performance is a
optimize the embedding based on dot-product similar- proxy, could be desirable.
ity. From this perspective, it is not surprising that the No single winner emerged from the comparisons
graph layouts by AROPE2, DW2, and GAE2 score of graph representation learning with graph layout

26 IEEE Computer Graphics and Applications May/June 2022

 APPLICATIONS

methods. Nonetheless, some general patterns did ACKNOWLEDGMENTS

emerge. This work was supported in part by the ERC under the
Graph layout methods: We found that FR and
EU’s 7th Framework and H2020 Programmes (ERC
FR-RTX resulted in readable layouts, score best on under Grants 615517 and 963924, in part by the Flem-
stress minimization, and achieved the highest link-pre- ish Government (“Onderzoeksprogramma Artificie €le
diction scores. The latter is especially surprising as the Intelligentie (AI) Vlaanderen”), and in part by the FWO
representation learning methods are focused on the (Projects 11J2322N, G0F9816N, and 3G042220).
generalization performance, through use of higher
order information (paths, convolutions), supervised
learning (use of negative edges), or both.
DRGRAPH is a fast method that preserves the REFERENCES
local graph neighborhoods better than both versions 1. B. Perozzi et al., “Deepwalk: Online learning of social
of FR. While the neighborhood preservation decreases representations,” in Proc. 20th ACM SIGKDD Int. Conf.
for larger graphs, it may be still the best choice to use Knowl. Discov Data Mining, 2014, pp. 701–710.
either of the computationally efficient graph layout 2. A. Tsitsulin et al., “Verse: Versatile graph embeddings
methods instead of a representation learning method. from similarity measures,” in Proc. World Wide Web
Graph representation learning methods: While the Conf., 2018, pp. 539–548.
native two-dimensional embeddings of CNE2 perform 3. J. F. Kruiger et al., “Graph layouts by t-SNE,” Comput.
well in terms of quality, it appears that the only reason Graphics Forum, vol. 36, no. 3, pp. 283–294, Jun. 2017.
to consider the AROPE2, DW2, and GAE2 embeddings 4. M. Zhu, W. Chen, Y. Hu, Y. Hou, L. Liu, and K. Zhang,
is to better understand the method itself. They are “DRgraph: An efficient graph layout algorithm for
outperformed, typically by a wide margin, by the t-SNE large-scale graphs by dimensionality reduction,” IEEE
embeddings of higher dimensional node representa- Trans. Vis. Comput. Graphics, vol. 27, no. 2, pp. 1666–
tions. DW128 is arguably the best representation learn- 1676, Feb. 2021.
ing method included in the evaluation. Compared to 5. O.-H. Kwon, T. Crnovrsanin, and K.-L. Ma, “What would a
DRGRAPH, it provides a more readable visualization graph look like in this layout? A machine learning
for the difficult Twitter graph, but is also 1000 times approach to large graph visualization,” IEEE Trans. Vis.
slower. For larger graphs, we expect the relative differ- Comput. Graphics, vol. 24, no. 1, pp. 478–488, Jan. 2018.
ence in runtime to grow further. 6. H. C. Purchase, “Metrics for graph drawing aesthetics,”
J. Vis. Lang. Comput., vol. 13, no. 5, pp. 501–516, 2002.
€ nger, “Large-graph layout
7. S. Hachul and M. Ju
Future Work algorithms at work: An experimental study,” J. Graph
Improving the visual quality: We observed that the Algorithms Appl., vol. 11, no. 2, pp. 345–369, 2007.
naive FR algorithm results in embeddings where 8. P. Eades et al., “Shape-based quality metrics for large
cluster structure is more pronounced compared to graph visualization,” J. Graph Algorithms Appl., vol. 21,
the grid optimized version where nodes are distrib- no. 1, pp. 29–53, 2017.
uted evenly across the layout. Adjusting the repul- 9. A. Mara et al., “EvalNE: A framework for evaluating
sive forces to depend on node degrees as in network embeddings on link prediction,” in Proc. EDML
ForceAtlas220 could improve the perceived cluster- (SDM), vol. 2436, 2019, pp. 5–13.
edness of the layout, by embedding leaf nodes 10. S. Zellmann et al., “Accelerating force-directed graph
closer to their parent. drawing with RT cores,” in Proc. IEEE Vis. Conf., 2020,
Combination of graph representation learning with pp. 96–100.
t-SNE: The t-SNE visualizations of the large Twitter 11. T. A. Davis and Y. Hu, “The university of florida sparse
graph (CNE16, DW128, and GAE16) are of high quality matrix collection,” ACM Trans. Math. Softw., vol. 38,
and suggest that preserving only the small distances no. 1, pp. 1–25, 2011.
of a low-dimensional node representation avoids the 12. J. Leskovec and J. McAuley, “Learning to discover
“hairball” embedding. It appears this allows deriving an social circles in ego networks,” Adv. Neural Inf.
integrated method that could outperform existing Process. Syst., vol. 25, pp. 539–547, 2012.
methods. Besides, it would be worthwhile to improve 13. Z. Zhang et al., “Arbitrary-order proximity preserved
the scalability of the DW128 combination. network embedding,” in Proc. 24th ACM SIGKDD Int.
Finally, we hope this evaluation inspires follow-up Conf. Knowl. Discov. Data Mining, 2018, pp. 2778–2786.
work to further connect the representation learning 14. B. Kang et al., “Conditional network embeddings,” in
with the graph drawing community. Proc. Int. Conf. Learn. Representations, 2019.

May/June 2022 IEEE Computer Graphics and Applications 27

APPLICATIONS

15. T. N. Kipf and M. Welling, “Variational graph auto- Germany. She is the corresponding author of this article. Con-
encoders,” in Proc. NeurIPS Workshop Bayesian Deep tact her at [email protected].
Learn., 2016.
16. T. Fruchterman and E. Reingold, “Graph drawing by
BO KANG is a postdoctoral researcher with the IDLab, Ghent
force-directed placement,” Softw.-Pract. Exp., vol. 21,
University, Ghent, Belgium. Kang received the Ph.D. degree in
no. 11, pp. 1129–1164, 1991.
computer science engineering from Ghent University, in 2019.
17. A. C. Mara, J. Lijffijt, and T. d. Bie, “Benchmarking
network embedding models for link prediction: Are we Contact him at [email protected].
making progress?,” in Proc. IEEE 7th Int. Conf. Data Sci.
Adv. Anal., 2020, pp. 138–147.
TIJL DE BIE is a full professor with Ghent University, Ghent,
18. L. Van der Maaten and G. Hinton, “Visualizing data
Belgium, having previously been affiliated with the KU
using t-SNE,” J. Mach. Learn. Res., vol. 9, no. 11,
pp. 2579–2605, 2008. Leuven, Leuven, Belgium, the University of Southampton,
19. A. Hagberg et al., “Exploring network structure, Southampton, U.K., University of California (UC) Berkeley,
dynamics, and function using networkx,” in Proc. SciPy, Berkeley, CA, USA, UC Davis, Davis, CA, and Bristol University,
2008, pp. 11–15. Bristol, U.K. Contact him at [email protected].
20. M. Jacomy et al., “ForceAtlas2, a continuous graph
layout algorithm for handy network visualization
designed for the gephi software,” PLoS One, vol. 9, JEFREY LIJFFIJT is a professor with Ghent University, Ghent,
no. 6, Jun. 2014, Art. no. e98679. Belgium. Lijffijt received the Doctor of Science degree in
technology from Aalto University, Espoo, Finland, in 2013.
EDITH HEITER is a Ph.D. student with the IDLab, Ghent Uni- Contact him at jefrey.lijffi[email protected].
versity, Ghent, Belgium. Heiter received the M.Sc. degree in
€ cken,
computer science from Saarland University, Saarbru Contact department editor Mike Potel at [email protected].

28 IEEE Computer Graphics and Applications May/June 2022