Chapter 2

Graph Representation Learning

Peng Cui, Lingfei Wu, Jian Pei, Liang Zhao and Xiao Wang

Abstract Graph representation learning aims at assigning nodes in a graph to low-

dimensional representations and effectively preserving the graph structures. Re-
cently, a significant amount of progress has been made toward this emerging graph
analysis paradigm. In this chapter, we first summarize the motivation of graph repre-
sentation learning. Afterwards and primarily, we provide a comprehensive overview
of a large number of graph representation learning methods in a systematic manner,
covering the traditional graph representation learning, modern graph representation
learning, and graph neural networks.

2.1 Graph Representation Learning: An Introduction

Many complex systems take the form of graphs, such as social networks, biological
networks, and information networks. It is well recognized that graph data is often
sophisticated and thus is challenging to deal with. To process graph data effectively,
the first critical challenge is to find effective graph data representation, that is, how
to represent graphs concisely so that advanced analytic tasks, such as pattern discov-
ery, analysis, and prediction, can be conducted efficiently in both time and space.

Liang Zhao
Department of Computer Science, Emory University, e-mail: [email protected]
Lingfei Wu
JD.COM Silicon Valley Research Center, e-mail: [email protected]
Peng Cui
Department of Computer Science, Tsinghua University, e-mail: [email protected]
Jian Pei
Department of Computer Science, Simon Fraser University, e-mail: [email protected]
Xiao Wang
Department of Computer Science, Beijing University of Posts and Telecommunications, e-mail:
[email protected]

17
18 Peng Cui, Lingfei Wu, Jian Pei, Liang Zhao and Xiao Wang

Traditionally, we usually represent a graph as G = (V , E ), where V is a node set

and E is an edge set. For large graphs, such as those with billions of nodes, the
traditional graph representation poses several challenges to graph processing and
analysis.
(1) High computational complexity. These relationships encoded by the edge
set E take most of the graph processing or analysis algorithms either iterative or
combinatorial computation steps. For example, a popular way is to use the shortest
or average path length between two nodes to represent their distance. To compute
such a distance using the traditional graph representation, we have to enumerate
many possible paths between two nodes, which is in nature a combinatorial prob-
lem. Such methods result in high computational complexity that prevents them from
being applicable to large-scale real-world graphs.
(2) Low parallelizability. Parallel and distributed computing is de facto to pro-
cess and analyze large-scale data. Graph data represented in the traditional way,
however, casts severe difficulties to design and implementat of parallel and dis-
tributed algorithms. The bottleneck is that nodes in a graph are coupled to each
other explicitly reflected by E. Thus, distributing different nodes in different shards
or servers often causes demandingly high communication cost among servers, and
holds back speed-up ratio.
(3) Inapplicability of machine learning methods. Recently, machine learning
methods, especially deep learning, are very powerful in many areas. For graph data
represented in the traditional way, however, most of the off-the-shelf machine learn-
ing methods may not be applicable. Those methods usually assume that data sam-
ples can be represented by independent vectors in a vector space, while the samples
in graph data (i.e., the nodes) are dependant to each other to some degree determined
by E. Although we can simply represent a node by its corresponding row vector in
the adjacency matrix of the graph, the extremely high dimensionality of such a rep-
resentation in a large graph with many nodes makes the in sequel graph processing
and analysis difficult.
To tackle these challenges, substantial effort has been committed to develop
novel graph representation learning, i.e., learning the dense and continuous low-
dimensional vector representations for nodes, so that the noise or redundant infor-
mation can be reduced and the intrinsic structure information can be preserved. In
the learned representation space, the relationships among the nodes, which were
originally represented by edges or other high-order topological measures in graphs,
are captured by the distances between nodes in the vector space, and the structural
characteristics of a node are encoded into its representation vector.
Basically, in order to make the representation space well supporting graph anal-
ysis tasks, there are two goals for graph representation learning. First, the original
graph can be reconstructed from the learned representation space. It requires that, if
there is an edge or relationship between two nodes, then the distance of these two
nodes in the representation space should be relatively small. Second, the learned rep-
resentation space can effectively support graph inference, such as predicting unseen
links, identifying important nodes, and inferring node labels. It should be noted that
a representation space with only the goal of graph reconstruction is not sufficient
2 Graph Representation Learning 19

for graph inference. After the representation is obtained, downstream tasks such as
node classification , node clustering , graph visualization and link prediction can be
dealt with based on these representations. Overall, there are three main categories of
graph representation learning methods: traditional graph embedding, modern graph
embedding, and graph neural networks, which will be introduced separately in the
following three sections.

2.2 Traditional Graph Embedding

Traditional graph embedding methods are originally studied as dimension reduction

techniques. A graph is usually constructed from a feature represented data set, like
image data set. As mentioned before, graph embedding usually has two goals, i.e.
reconstructing original graph structures and support graph inference. The objective
functions of traditional graph embedding methods mainly target the goal of graph
reconstruction.
Specifically, Tenenbaum et al (2000) first constructs a neighborhood graph G us-
ing connectivity algorithms such as K nearest neighbors (KNN). Then based on G,
the shortest path between different data can be computed. Consequently, for all the
N data entries in the data set, we have the matrix of graph distances. Finally, the
classical multidimensional scaling (MDS) method is applied to the matrix to obtain
the coordinate vectors. The representations learned by Isomap approximately pre-
serve the geodesic distances of the entry pairs in the low-dimensional space. The key
problem of Isomap is its high complexity due to the computing of pair-wise short-
est pathes. Locally linear embedding (LLE) (Roweis and Saul, 2000) is proposed
to eliminate the need to estimate the pairwise distances between widely separated
entries. LLE assumes that each entry and its neighbors lie on or close to a locally
linear patch of a mainfold. To characterize the local geometry, each entry can be
reconstructed from its neighbors. Finally, in the low-dimensional space, LLE con-
structs a neighborhood-preserving mapping based on locally linear reconstruction.
Laplacian eigenmaps (LE) (Belkin and Niyogi, 2002) also begins with construct-
ing a graph using e-neighborhoods or K nearest neighbors. Then the heat kernel
(Berline et al, 2003) is utilized to choose the weight of two nodes in the graph. Fi-
nally, the node representations can be obtained by based on the Laplacian matrix
regularization. Furthermore, the locality preserving projection (LPP) (Berline et al,
2003), a linear approximation of the nonlinear LE, is proposed.
These methods are extended in the rich literature of graph embedding by consid-
ering different characteristics of the constructed graphs (Fu and Ma, 2012). We can
find that traditional graph embedding mostly works on graphs constructed from fea-
ture represented data sets, where the proximity among nodes encoded by the edge
weights is well defined in the original feature space. While, in contrast, modern
graph embedding, which will be introduced in the following, mostly works on natu-
rally formed networks, such as social networks, biology networks, and e-commerce
networks. In those networks, the proximities among nodes are not explicitly or di-
20 Peng Cui, Lingfei Wu, Jian Pei, Liang Zhao and Xiao Wang

rectly defined. For example, an edge between two nodes usually just implies there is
a relationship between them, but cannot indicate the specific proximity. Also, even
if there is no edge between two nodes, we cannot say the proximity between these
two nodes is zero. The definition of node proximities depends on specific analytic
tasks and application scenarios. Therefore, modern graph embedding usually incor-
porates rich information, such as network structures, properties, side information
and advanced information, to facilitate different problems and applications. Modern
graph embedding needs to target both of goals mentioned before. In view of this,
traditional graph embedding can be regarded as a special case of modern graph em-
bedding, and the recent research progress on modern graph embedding pays more
attention to network inference.

2.3 Modern Graph Embedding

To well support network inference, modern graph embedding considers much richer
information in a graph. According to the types of information that are preserved in
graph representation learning, the existing methods can be categorized into three
categories: (1) graph structures and properties preserving graph embedding, (2)
graph representation learning with side information and (3) advanced information
preserving graph representation learning. In technique view, different models are
adopted to incorporate different types of information or address different goals. The
commonly used models include matrix factorization, random walk, deep neural net-
works and their variations.

2.3.1 Structure-Property Preserving Graph Representation

Learning

Among all the information encoded in a graph, graph structures and properties are
two crucial factors that largely affect graph inference. Thus, one basic requirement
of graph representation learning is to appropriately preserve graph structures and
capture properties of graphs. Often, graph structures include first-order structures
and higher-order structures, such as second-order structures and community struc-
tures. Graphs with different types have different properties. For example, directed
graphs have the asymmetric transitivity property. The structural balance theory is
widely applicable to signed graphs.

2.3.1.1 Structure Preserving Graph Representation Learning

Graph structures can be categorized into different groups that present at differ-
ent granularities. The commonly exploited graph structures in graph representation
2 Graph Representation Learning 21

learning include neighborhood structure, high-order node proximity and graph com-
munities.
How to define the neighborhood structure in a graph is the first challenge. Based
on the discovery that the distribution of nodes appearing in short random walks is
similar to the distribution of words in natural language, DeepWalk (Perozzi et al,
2014) employs the random walks to capture the neighborhood structure. Then for
each walk sequence generated by random walks, following Skip-Gram, DeepWalk
aims to maximize the probability of the neighbors of a node in a walk sequence.
Node2vec defines a flexible notion of a node’s graph neighborhood and designs
a second order random walks strategy to sample the neighborhood nodes, which
can smoothly interpolate between breadth-first sampling (BFS) and depth-first sam-
pling (DFS). Besides the neighborhood structure, LINE (Tang et al, 2015b) is pro-
posed for large scale network embedding, which can preserve the first and second
order proximities. The first order proximity is the observed pairwise proximity be-
tween two nodes. The second order proximity is determined by the similarity of
the “contexts” (neighbors) of two nodes. Both are important in measuring the re-
lationships between two nodes. Essentially, LINE is based on the shallow model,
consequently, the representation ability is limited. SDNE (Wang et al, 2016) pro-
poses a deep model for network embedding, which also aims at capturing the first
and second order proximites. SDNE uses the deep auto-encoder architecture with
multiple non-linear layers to preserve the second order proximity. To preserve the
first-order proximity, the idea of Laplacian eigenmaps (Belkin and Niyogi, 2002)
is adopted. Wang et al (2017g) propose a modularized nonnegative matrix factor-
ization (M-NMF) model for graph representation learning, which aims to preserve
both the microscopic structure, i.e., the first-order and second-order proximities of
nodes, and the mesoscopic community structure (Girvan and Newman, 2002). They
adopt the NMF model (Févotte and Idier, 2011) to preserve the microscopic struc-
ture. Meanwhile, the community structure is detected by modularity maximization
(Newman, 2006a). Then, they introduce an auxiliary community representation ma-
trix to bridge the representations of nodes with the community structure. In this
way, the learned representations of nodes are constrained by both the microscopic
structure and community structure.
In summary, many network embedding methods aim to preserve the local struc-
ture of a node, including neighborhood structure, high-order proximity as well as
community structure, in the latent low-dimensional space. Both linear and non-
linear models are attempted, demonstrating the large potential of deep models in
network embedding.

2.3.1.2 Property Preserving Graph Representation Learning

Currently, most of the existing property preserving graph representation learning

methods focus on graph transitivity in all types of graphs and the structural balance
property in signed graphs.
22 Peng Cui, Lingfei Wu, Jian Pei, Liang Zhao and Xiao Wang

We usually demonstrate that the transitivity usually exists in a graph. But mean-
while, we can find that preserving such a property is not challenging, because in a
metric space, the distance between different data points naturally satisfies the trian-
gle inequality. However, this is not always true in the real world. Ou et al (2015) aim
to preserve the non-transitivity property via latent similarity components. The non-
transitivity property declares that, for nodes v1 , v2 and v3 in a graph where (v1 ; v2 )
and (v2 ; v3 ) are similar pairs, (v1 ; v3 ) may be a dissimilar pair. For example, in a
social network, a student may connect with his classmates and his family, while his
classmates and family are probably very different. The main idea is that they learn
multiple node embeddings, and then compare different nodes based on multiple
similarities, rather than one similarity. They observe that if two nodes have a large
semantic similarity, at least one of the structure similarities is large, otherwise, all
of the similarities are small. In a directed graph, it usually has the asymmetric tran-
sitivity property. Asymmetric transitivity indicates that, if there is a directed edge
from node i to node j and a directed edge from j to v, there is likely a directed edge
from i to v, but not from v to i. In order to measure this high-order proximity, HOPE
(Ou et al, 2016) summarizes four measurements in a general formulation, and then
utilizes a generalized SVD problem to factorize the high-order proximity (Paige and
Saunders, 1981), such that the time complexity of HOPE is largely reduced, which
means HOPE is scalable for large scale networks. In a signed graph with both of
positive and negative edges, the social theories, such as structural balance theory
(Cartwright and Harary, 1956; Cygan et al, 2012), which are very different from the
unsigned graph. The structural balance theory demonstrates that users in a signed
social network should be able to have their “friends” closer than their “foes”. To
model the structural balance phenomenon, SiNE (Wang et al, 2017f) utilizes a deep
learning model consisting of two deep graphs with non-linear functions.
The importance of maintaining network properties in network embedding space,
especially the properties that largely affect the evolution and formation of networks,
has been well recognized. The key challenge is how to address the disparity and het-
erogeneity of the original network space and the embedding vector space at property
level. Generally, most of the structure and property preserving methods take high
order proximities of nodes into account, which demonstrate the importance of pre-
serving high order structures in network embedding. The difference is the strategy
of obtaining the high order structures. Some methods implicitly preserve highorder
structure by assuming a generative mechanism from a node to its neighbors, while
some other methods realize this by explicitly approximating high-order proximities
in the embedding space. As topology structures are the most notable characteristic
of networks, structure-preserving network methods embody a large part of the lit-
erature. Comparatively, property preserving network embedding is a relatively new
research topic and is only studied lightly. As network properties usually drive the
formation and evolution of networks, it shows great potential for future research and
applications.
2 Graph Representation Learning 23

2.3.2 Graph Representation Learning with Side Information

Besides graph structures, side information is another important information source

for graph representation learning. Side information in the context of graph represen-
tation learning can be divided into two categories: node content and types of nodes
and edges. Their difference is the way of integrating network structures and side
information.
Graph Representation Learning with Node Content. In some types of graphs,
like information networks, nodes are acompanied with rich information, such as
node labels, attributes or even semantic descriptions. How to combine them with
the network topology in graph representation learning arouses considerable research
interests. Tu et al (2016) propose a semi-supervised graph embedding algorithm,
MMDW, by leveraging labeling information of nodes. MMDW is also based on the
DeepWalk-derived matrix factorization. MMDW adopts support vector machines
(SVM) (Hearst et al, 1998) and incorporates the label information to find an optimal
classifying boundary. Yang et al (2015b) propose TADW that takes the rich informa-
tion (e.g., text) associated with nodes into account when they learn the low dimen-
sional representations of nodes. Pan et al (2016) propose a coupled deep model that
incorporates graph structures, node attributes and node labels into graph embedding.
Although different methods adopt different strategies to integrate node content and
network topology, they all assume that node content provides additional proximity
information to constrain the representations of nodes.
Heterogeneous Graph Representation Learning. Different from graphs with node
content, heterogeneous graphs consist of different types of nodes and links. How to
unify the heterogeneous types of nodes and links in graph embedding is also an
interesting and challenging problem. Jacob et al (2014) propose a heterogeneous
social graph representation learning algorithm for classifying nodes. They learn the
representations of all types of nodes in a common vector space, and perform the
inference in this space. Chang et al (2015) propose a deep graph representation
learning algorithm for heterogeneous graphs, whose nodes have various types(e.g.,
images and texts). The nonlinear embeddings of images and texts are learned by
a CNN model and the fully connected layers, respectively. Huang and Mamoulis
(2017) propose a meta path similarity preserving heterogeneous information graph
representation learning algorithm. To model a particular relationship, a meta path
(Sun et al, 2011) is a sequence of object types with edge types in between.
In the methods preserving side information, side information introduces addi-
tional proximity measures so that the relationships between nodes can be learned
more comprehensively. Their difference is the way of integrating network struc-
turess and side information. Many of them are naturally extensions from structure
preserving network embedding methods.
24 Peng Cui, Lingfei Wu, Jian Pei, Liang Zhao and Xiao Wang

2.3.3 Advanced Information Preserving Graph Representation

Learning

Different from side information, the advanced information refers to the supervised
or pseudo supervised information in a specific task. The advanced information pre-
serving network embedding usually consists of two parts. One is to preserve the
network structure so as to learn the representations of nodes. The other is to estab-
lish the connection between the representations of nodes and the target task. The
combination of advanced information and network embedding techniques enables
representation learning for networks.
Information Diffusion. Information diffusion (Guille et al, 2013) is an ubiquitous
phenomenon on the web, especially in social networks. Bourigault et al (2014) pro-
pose a graph representation learning algorithm for predicting information diffusion
in social network. The goal of the proposed algorithm is to learn the representations
of nodes in the latent space such that the diffusion kernel can best explain the cas-
cades in the training set. The basic idea is to map the observed information diffusion
process into a heat diffusion process modeled by a diffusion kernel in the continu-
ous space. The kernel describes that the closer a node in the latent space is from
the source node, the sooner it is infected by information from the source node. The
cascade prediction problem here is defined as predicting the increment of cascade
size after a given time interval (Li et al, 2017a). Li et al (2017a) argue that the pre-
vious work on cascade prediction all depends on the bag of hand-crafting features
to represent the cascade and graph structures. Instead, they present an end-to-end
deep learning model to solve this problem using the idea of graph embedding. The
whole procedure is able to learn the representation of cascade graph in an end-to-end
manner.
Anomaly Detection. Anomaly detection has been widely investigated in previous
work (Akoglu et al, 2015). Anomaly detection in graphs aims to infer the structural
inconsistencies, which means the anomalous nodes that connect to various diverse
influential communities (Hu et al, 2016), (Burt, 2004). Hu et al (2016) propose a
graph embedding based method for anomaly detection. They assume that the com-
munity memberships of two linked nodes should be similar. An anomaly node is
one connecting to a set of different communities. Since the learned embedding of
nodes captures the correlations between nodes and communities, based on the em-
bedding, they propose a new measure to indicate the anomalousness level of a node.
The larger the value of the measure, the higher the propensity for a node being an
anomaly node.
Graph Alignment. The goal of graph alignment is to establish the correspon-
dence between the nodes from two graphs, i.e., to predict the anchor links across
two graphs. The same users who are shared by different social networks naturally
form the anchor links, and these links bridge the different graphs. The anchor link
prediction problem is, given a source graph,a target graph and a set of observed
anchor links, to identify the hidden anchor links across the two graphs. Man et al
(2016) propose a graph representation learning algorithm to solve this problem. The
2 Graph Representation Learning 25

learned representations can preserve the graph structures and respect the observed
anchor links.
Advanced information preserving graph embedding usually consists of two parts.
One is to preserve the graph structures so as to learn the representations of nodes.
The other is to establish the connection between the representations of nodes and the
target task. The first one is similar to structure and property preserving network em-
bedding, while the second one usually needs to consider the domain knowledge of a
specific task. The domain knowledge encoded by the advanced information makes
it possible to develop end-to-end solutions for network applications. Compared with
the hand-crafted network features, such as numerous network centrality measures,
the combination of advanced information and network embedding techniques en-
ables representation learning for networks. Many network applications may be ben-
efitted from this new paradigm.

2.4 Graph Neural Networks

Over the past decade, deep learning has become the “crown jewel” of artificial intel-
ligence and machine learning, showing superior performance in acoustics, images
and natural language processing, etc. Although it is well known that graphs are ubiq-
uitous in the real world, it is very challenging to utilize deep learning methods to
analyze graph data. This problem is non-trivial because of the following challenges:
(1) Irregular structures of graphs. Unlike images, audio, and text, which have a clear
grid structure, graphs have irregular structures, making it hard to generalize some
of the basic mathematical operations to graphs. For example, defining convolution
and pooling operations, which are the fundamental operations in convolutional neu-
ral networks (CNNs), for graph data is not straightforward. (2) Heterogeneity and
diversity of graphs. A graph itself can be complicated, containing diverse types and
properties. These diverse types, properties, and tasks require different model archi-
tectures to tackle specific problems. (3) Large-scale graphs. In the big-data era, real
graphs can easily have millions or billions of nodes and edges. How to design scal-
able models, preferably models that have a linear time complexity with respect to the
graph size, is a key problem. (4) Incorporating interdisciplinary knowledge. Graphs
are often connected to other disciplines, such as biology, chemistry, and social sci-
ences. This interdisciplinary nature provides both opportunities and challenges: do-
main knowledge can be leveraged to solve specific problems but integrating domain
knowledge can complicate model designs.
Currently, graph neural networks have attracted considerable research attention
over the past several years. The adopted architectures and training strategies vary
greatly, ranging from supervised to unsupervised and from convolutional to re-
cursive, including graph recurrent neural networks (Graph RNNs), graph convo-
lutional networks (GCNs), graph autoencoders (GAEs), graph reinforcement learn-
ing (Graph RL), and graph adversarial methods. Specifically, Graroperty h RNNs
capture recursive and sequential patterns of graphs by modeling states at either the
26 Peng Cui, Lingfei Wu, Jian Pei, Liang Zhao and Xiao Wang

node-level or the graph-level; GCNs define convolution and readout operations on

irregular graph structures to capture common local and global structural patterns;
GAEs assume low-rank graph structures and adopt unsupervised methods for node
representation learning; Graph RL defines graph-based actions and rewards to ob-
tain feedbacks on graph tasks while following constraints; Graph adversarial meth-
ods adopt adversarial training techniques to enhance the generalization ability of
graphbased models and test their robustness by adversarial attacks.
There are many ongoing or future research directions which are also worthy of
further study, including new models for unstudied graph structures, compositional-
ity of existing models, dynamic graphs, interpretability and robustness, etc. On the
whole, deep learning on graphs is a promising and fast-developing research field
that both offers exciting opportunities and presents many challenges. Studying deep
learning on graphs constitutes a critical building block in modeling relational data,
and it is an important step towards a future with better machine learning and artifi-
cial intelligence techniques.

2.5 Summary

In this chapter, we introduce the motivation of graph representation learning. Then

in Section 2, we discuss the traditional graph embedding methods and the mod-
ern graph embedding methods are introduced in Section 3. Basically, the structure
and property preserving graph representation learning is the foundation. If one can-
not preserve well the graph structures and retain the important graph properties in
the representation space, serious information will be lost, which hurts the analytic
tasks in sequel. Based on the structures and property preserving graph representation
learning, one may apply the off-the-shelf machine learning methods. If some side
information is available, it can be incorporated into graph representation learning.
Furthermore, the domain knowledge of some certain applications as advanced infor-
mation can be considered. As shown in Section 4, utilizing deep learning methods
on graphs is a promising and fast-developing research field that both offers excit-
ing opportunities and presents many challenges. Studying deep learning on graphs
constitutes a critical building block in modeling relational data, and it is an impor-
tant step towards a future with better machine learning and artificial intelligence
techniques.