Journal of Data Science 20 (2), 111–134 DOI: 10.

6339/22-JDS1047
April 2022 Data Science Reviews

A Review on Graph Neural Network Methods in Financial

Applications

Jianian Wang1 , Sheng Zhang1 , Yanghua Xiao2,∗ , and Rui Song1,∗

1 Department of Statistics, North Carolina State University, Raleigh, United States
2 School of Computer Science, Fudan University, Shanghai, China

Abstract
With multiple components and relations, financial data are often presented as graph data, since it
could represent both the individual features and the complicated relations. Due to the complexity
and volatility of the financial market, the graph constructed on the financial data is often
heterogeneous or time-varying, which imposes challenges on modeling technology. Among the
graph modeling technologies, graph neural network (GNN) models are able to handle the complex
graph structure and achieve great performance and thus could be used to solve financial tasks.
In this work, we provide a comprehensive review of GNN models in recent financial context. We
first categorize the commonly-used financial graphs and summarize the feature processing step
for each node. Then we summarize the GNN methodology for each graph type, application in
each area, and propose some potential research areas.
Keywords deep learning; finance; graph convolutional network; graph representation learning

1 Introduction
As the data collection techniques grow, graph data are commonly collected in many areas in-
cluding social sciences, transportation systems, chemistry, and physics (Wu et al., 2020). Repre-
senting complex relational data, graph data contain both the individual node information and
the structural information. Recently, there are growing interests in developing machine-learning
methods to model the graph data of various domains. Among them, graph neural network (GNN)
methods could achieve great performance on various tasks, including node classification, edge
prediction, and graph classification (Kipf and Welling, 2017; Zhang and Chen, 2018; Xu et al.,
2018). Performing node aggregation and updates, graph neural network models extend the deep
learning methodology to graphs and are gaining popularity.
A financial system is a complex system with many components and sophisticated relations,
which may be frequently updated. To represent the relational data in the financial domain,
graphs are commonly constructed, including the transaction network (Weber et al., 2019), user-
item review graph (Dou et al., 2020), and stock relation graph (Feng et al., 2019). By converting
the financial task into a node classification task, GNN methods are commonly utilized since it
performs well among graph modeling methods (Liu et al., 2019). For instance, GNN could be
utilized in a stock prediction task, by formulating it as a node classification task, where each
node represents a stock and edges represent relations between companies. Figure 1 demonstrates
the workflow of a stock prediction task using GNN methods. However, the complex nature

∗ Corresponding author. Email: [email protected] or [email protected].

© 2022 The Author(s). Published by the School of Statistics and the Center for Applied Statistics, Renmin
University of China. Open access article under the CC BY license.
Received November 3, 2021; Accepted April 10, 2022
112 Wang, J. et al.

Figure 1: Workflow for stock movement prediction task using GNN methodology. The graph
construction and feature processing steps present stock information in a graph and a feature
matrix, which is then used as the input for the GNN model. In the graph, nodes are connected
if there exist some relationships between stocks, such as supplier, competitor, shake-holder, etc.
A multi-layer perception layer (MLP) is used to output the price prediction result.

of financial systems may result in multiple data sources and complicated graph structures,
which imposes challenges on feature processing, graph construction, and graph neural network
modeling. Represented as numerical sequences or textual information, financial data need to be
processed with caution to keep the temporal pattern or semantic meanings. Also, the multi-facet
nature of financial relations make it hard to construct a graph to capture the relations. Moreover,
the financial-related graph is often heterogeneous or time-varying, which impose challenges on
existing graph neural network models. What’s more, to reflect some financial patterns (e.g.
device aggregation pattern, see Section 5.4 for details), GNN methods may need to be modified
such as changing losses and adding additional layers. Since financial systems process unique
characteristics and receive great attention, it is of significant importance to discuss and summary
the GNN methodology developed for financial tasks.
There are several recent reviews on graph neural networks. Among them, Wu et al. (2020)
present a comprehensive review on graph neural networks and categorize the GNNs into four
categories: recurrent graph neural networks, convolutional graph neural networks, graph auto-
encoders, and spatial-temporal graph neural networks. Zhou et al. (2020) provide a taxonomy
on GNN models based on graph type, training methods, and propagation steps. There is also
literature focusing on limited types of GNNs. Zhang et al. (2019) focus on graph convolutional
networks (GCN) and introduce two taxonomies to group the existing GCNs. Lee et al. (2019)
survey the literature on graph attention models and provided detailed examples on each type
of method. However, the aforementioned reviews focus on the general methodology and provide
little details for applications, seldom mentioning the financial application. Without covering
GNN models developed based on financial contexts, the reviewed models may not be applicable
to financial tasks due to the complexity of financial data. On the other side, review papers
focusing on the financial domain haven’t covered GNN methodologies in detail yet. Ozbayoglu
et al. (2020) summarize the machine learning and deep learning models in the financial field,
without mentioning the GNN methodologies. Huang et al. (2020) survey the financial deep
learning models in the finance and bank industry, and the GNN models are not covered. Jiang
(2021) review stock prediction-related machine-learning mythologies and mention GNN models
very briefly. In summary, existing GNN surveys focus on modeling methodology and do not
emphasize the financial application of GNN methods, while surveys on financial applications
 A Review on GNN in Finance 113

don’t cover the GNN models in detail. To fill the gap, in this survey, we provide a systematic
and comprehensive review of graph neural network methods in the financial application.
In this paper, we present a thorough survey on graph neural network models with financial
application. We provide a comprehensive review of graph neural networks and summarize the
corresponding methods. This survey has contributions as follows.
• We systemically categorize the commonly-used financial graphs based on graph charac-
teristics and provide a thorough list of graphs. Graphs are categorized into five groups:
homogeneous graph, directed graph, bipartite graph, multi-relation graph, and dynamic
graph. We also present the GNN models according to their graph types, so that this review
could serve as a guide for implementing GNNs on real-life datasets.
• We provide a comprehensive list of financial applications that GNN methods are ap-
plied. We categorized the applications into five categories: stock movement prediction,
loan default risk prediction, recommender system of e-commerce, fraud detection, and
event prediction.
• We summarize various aspects of information for each application, including features,
graphs, GNN models, and available codes. A GitHub1 page is built to document the
collection of information. This work could be considered as a resource to understand,
implement and develop GNN models on multiple financial tasks.
• We identify five challenges and discuss the recent progress. We also suggest future direc-
tions for these problems.
The rest of the paper is organized as follows. Section 2 classifies financial graphs into
different categories based on its characteristics. Section 3 summarizes the commonly-used feature
processing techniques for each node in the graph. Section 4 presents the GNN methodology used
for each graph type. Section 5 provides a collection of application areas. Section 6 proposes some
challenges that could be future directions of research.

2 Graph Categorization
When preparing the data, how to construct the graph to represent the structural information
is essential and the type for the constructed graph could determine the follow-up modeling
methodology. In this section, we present the categorization of the graph based on its construction
methods and graph types. Table 1 presents a comprehensive list of graphs for financial tasks.

2.1 Graph-Related Definition

In this section, we provide some graph-related definitions for better understanding of this article.

Definition 2.1 (Graph). A graph G is defined by a pair: G = (V, E), where V = {v1 , . . . , vn } is
a set of n nodes and E is a set of edges, where eij = (vi , vj ) ∈ E denotes an edge joining node vi
and node vj .

Definition 2.2 (Adjacency matrix). An adjacency matrix A is an n × n matrix, where Aij

represents the connection status between node vi and node vj . For an unweighted graph, the
adjacency matrix could be an binary matrix where Aij = 1 if eij ∈ E and 0 otherwise.

1 Github link: https://github.com/jackieD14/Graph-models-in-finance-application

114 Wang, J. et al.

Table 1: Summary of financial-related graphs.

Graph Construction Graph Application 45 Reference

method 2 type 3
Sector-industry stock relation Knowledge Multi Stock Feng et al. (2019)
network
Wiki company-based relation Knowledge Multi Stock Feng et al. (2019);
network Sawhney et al.
(2020a); Ying et al.
(2020)
Supplier, customer, partner and Knowledge Multi Stock Matsunaga et al.
shareholder relation graph (2019)
Corporation shareholder net- Knowledge Homo Stock Chen et al. (2018)
work
Stock correlation graph Similarity Multi Stock Li et al. (2020a)
Stock earning call graph Knowledge Bipartite Stock Sawhney et al.
(2020b)
News co-occurrence graph Similarity Homo Stock Liou et al. (2021)
User-relation graph Data Homo Loan Wang et al. (2019)
User-app graph Data Bipartite Loan Wang et al. (2019)
User-nickname graph Data Bipartite Loan Wang et al. (2019)
User-address graph Data Bipartite Loan Wang et al. (2019)
Guarantee network Data Directed Loan Cheng et al. (2019,
2020)
Temporal guarantee network Data Dynamic Loan Cheng and Zhang L
Wang X (2020b)
Temporal small business en- Data Dynamic Loan Yang et al. (2020)
trepreneur network
Alipay user and applet graph Data Dynamic Loan Hu et al. (2020)
User relationship graph Data Multi Loan Liang et al. (2021)
Auto relation network Data Multi Loan Xu et al. (2021)
Loan application event graph Data Directed Loan Harl et al. (2020)
Borrower relations’ network Similarity Directed Loan Lee et al. (2021)
Xianyu comment graph Similarity Homo E-comm Li et al. (2019a)
Yelp review network Data Bipartite E-comm Zhang et al. (2020);
Dou et al. (2020)
Amazon review network Data Bipartite E-comm Zhang et al. (2020);
Dou et al. (2020);
Kudo et al. (2020)
E-commerce user-item network Data Bipartite E-comm Li et al. (2019b)
Taobao user-item network Data Bipartite E-comm Li et al. (2020c)
Device sharing graph Data Bipartite Fraud Liang et al. (2019)
JD Finance anti-fraud graph Data Multi Fraud Lv et al. (2019)
Transaction records graph Data Multi Fraud Rao et al. (2020)
Iqiyi user network Data Multi Fraud Zhu et al. (2020)
CMU simulated user activity Similarity Homo Fraud Jiang et al. (2019)
network
Alipay one-month account- Data Bipartite Fraud Liu et al. (2018)
device network
Alipay one-week account-device Data Bipartite Fraud Liu et al. (2019)
network
Account-registration graph Data Dynamic Fraud Rao et al. (2020)
Bitcoin-alpha graph Data Directed Fraud Zhao et al. (2021)
 A Review on GNN in Finance 115

Definition 2.3 (Undirected graph and Directed graph). A undirected graph is a graph where
the edges are undirected. A directed graph is a graph where the edges have orientations. eij =
(vi , vj ) ∈ E denotes an edge pointing from node vi to node vj .
Remark. Undirected graph has a symmetric adjacency matrix, i.e., Aij = Aj i .
Definition 2.4 (Bipartite graph). A Bipartite graph is a graph whose nodes could be divided
into two non-empty and disjoint sets U, W, such that every edge connects a node in U and a
node in W.
Definition 2.5 (Homogeneous graph and Heterogeneous graph). In a graph G = (V, E), we can
assign a type to each node and edge; in this case, the graph is denoted as G = (V, E, A, R),
where each node vi ∈ V is associated with its type ai ∈ A, and each edge eij ∈ E is associated
with its type rij ∈ R. A homogeneous graph is a graph whose nodes are of the same type and
edges are of the same type. Otherwise, the graph is heterogeneous.
Definition 2.6 (Multi-relation graph). A Multi-relation graph is a graph where edges have
different types.
Definition 2.7 (Dynamic graph). A dynamic graph is defined as a sequence of graphs G seq =
{G1 , . . . , GT }, where Gi = (Vi , Ei ), for i = 1, . . . , T , where Vi , Ei are the set of nodes and edges
for ith graph in the sequence respectively.

2.2 Graph Categorization by Construction Methods

In this section, we elaborate on frequently-used graph construction methods so that researchers
could better understand, choose and construct graph data.

2.2.1 Data-Based Construction

Some types of data could be naturally represented as a graph since they contain relations among
data objects. For instance, Wang et al. (2019) construct a user-relationship graph, where users
are linked by an edge if they are labeled classmates, friends, or workmates in the data. Liu et al.
(2018) construct an account-device network, where nodes are either accounts or devices. Edges
connect an account node to a device node if the account has activities on the device. This type
of construction method is based on the nature of data and could be used on data where relations
are clearly defined.

2.2.2 Knowledge-Based Construction

Sometimes, data may not contain relational information, but relations could be found in knowl-
edge bases. A knowledge base is a collection of descriptive data and contains numerous entities
and their relations. For example, Wikidata (Vrandečić and Krötzsch, 2014) is one of the largest
open-domain knowledge bases which provides support for Wikipedia, an online encyclopedia. A
graph could be built utilizing the relations extracted from knowledge bases. In order to predict
stock movement, Feng et al. (2019) extract company relations from Wikidata, such as supplier,
provider, partner, etc, and constructed a company-based relation network. This type of graph-
construction method brings new information into the graph by utilizing the knowledge bases,
which may improve modeling performance. However, it may take some effort to process the
complicated data structure and the massive amount of information of knowledge bases.
116 Wang, J. et al.

Figure 2: Graph categorization based on graph characteristics. Each color of the circle represents
a node type and each color of the line represents an edge type. Arrows represent directed edges.
A homogeneous graph is a graph with one type of node and one type of edge. A directed graph is
a graph with directed edges. A bipartite graph is a graph with two types of nodes and edges only
exist between nodes of different types. A multi-relation graph has edges with different types. A
dynamic graph is a sequence of graphs.

2.2.3 Similarity-Based Construction

There also exist cases that neither the data nor knowledge bases contain relations, but there
may be some hidden relationships in the data. To mine the underlying relationship, a commonly-
used approach is to calculate a similarity measure of the features for different observations and
construct relations if the similarity value is greater than a threshold. For instance, Li et al.
(2020a) construct a stock correlation graph based on the cosine similarity of stocks’ historic
market price. Two stocks are then connected if the absolute value of their correlation is larger
than a threshold. This type of construction method could be easily implemented and understood.
However, in this type of construction, feature information is represented both in the graph
adjacency matrix and the feature matrix. The overlapping information may lead to doubts that
whether the graph representation is still necessary. Also, how to set the threshold is an issue,
and justification may be needed for the selected similarity threshold value.

2.3 Graph Categorization by Graph Types

In this section, we categorize graphs into five categories based on their characteristics and provide
examples in the financial context. Since different types of graphs may impose various challenges
on modeling technology, we discuss GNN methods for each graph type in section 4 to pro-
vide solutions respectively. Figure 2 provides visualization for each graph type: homogeneous
graph, directed graph, bipartite graph, multi-relation graph, and dynamic graph. It is also worth
mentioning that a graph may be categorized into multiple types.
 A Review on GNN in Finance 117

2.3.1 Homogeneous Graph

A homogeneous graph is a graph with one type of node and one type of edge. For instance, Liou
et al. (2021) construct a financial news co-occurrence graph, where two companies are connected
if they are tagged in the same news articles. Li et al. (2020b) build a transaction network, where
nodes are accounts and are connected when there exist transactions between them. This type of
graph has a relatively simple structure and the majority of the GNN methods could be applied
to model this type of graph.

2.3.2 Directed Graph

A directed graph is a graph where edges have orientations. For instance, Cheng et al. (2019)
construct a guarantee network where nodes are the companies and edges represent the guarantee
relationship. Since the guarantor has the obligation to pay the debt for the borrower, but not the
other way round, this type of guarantee relationship is one-sided and could be represented in a
directed edge. In general, a directed graph may have an asymmetric adjacency matrix and thus
cannot be semi-definite. Since some GNN mythologies are developed for semi-definite adjacency
matrices, they may not be suitable for the directed graph.

2.3.3 Bipartite Graph

A bipartite graph is a graph with two types of nodes and edges only exist between nodes of
different types. For instance, Liu et al. (2018) construct an account-device network in a fraud
detection task. Nodes could be either accounts or devices, with edges connecting them if the
account has activities on the device. Li et al. (2019a) extract a user-item network using the
rating data, where nodes could be either a user or an item. An edge exists if a user has rated
the item. A bipartite graph is commonly used, when data can be divided into two groups and
the interaction between two groups matters. A bipartite graph could be seen as a specific case
of a multi-relation graph discussed in the next session and GNN methods developed for the
multi-relation graph are applicable on a bipartite graph as well.

2.3.4 Multi-Relation Graph

Sometimes, edges may have multiple types to represent different relations between nodes. For
instance, Wang et al. (2019) construct a user-relation graph, where nodes are the users of an
e-commerce platform. There are multiple edge types representing various relationships including
friendship, workmates and classmates. Dou et al. (2020) build a review graph where users are
represented as nodes in the graph. Three types of relations between users are defined to capture
their behavioral patterns: reviewing the same product, having the same star rating, and having
similar texts. With multiple edge types, this type of graph contains more information about the
relationship between nodes, and thus how to capture this information is critical when developing
the GNN methodology.

2.3.5 Dynamic Graph

A dynamic graph is a sequence of graphs, where each graph could have an adjacency matrix
and feature matrix. For example, Cheng and Zhang L Wang X (2020b) construct a temporal
guarantee network, representing the guarantee relationship in each time step. This type of graph
is commonly used to represent the changes in both relations and features, as time goes. Since
118 Wang, J. et al.

Figure 3: Feature processing for sequential features and textual information. For sequential
numerical features, recurrent neural network (RNN) based approaches are commonly used to
capture the temporal dependencies. For text features, it is often processed utilizing natural lan-
guage processing (NLP) methods including word embedding, sentence embedding, and language
models, to convert the unstructured data to structured ones.

both nodes and edges could appear and disappear, it is hard to perform some graph operations
that require fixed dimensions of matrices. Thus, capturing the dynamically of the graph is
challenging and requires a more sophisticated methodology.

3 Feature Processing
With diverse data sources in the financial field, node features are commonly formatted as se-
quential numerical features or textual information. These data formats impose challenges on the
feature processing step since GNN methods could not be directly applied to these data formats.
In this section, we summarize the commonly-used feature processing techniques and how they
solve these challenges.

3.1 Sequential Numerical Data

Updating information as time goes, the financial industry has a rich source of time-series data.
Indexed using timestamps, features could be seen as sequential which requires appropriate mod-
eling. Consider the feature matrix at time s, Xs ∈ Rn×p×l , where n is the number of nodes, p
is the dimension of features at each time point, l is the length of the sequence. For example,
in a stock prediction task, we have a stock relation graph, with n stocks as nodes. The feature
matrix X s represents the p features in the past l days from time s for these n stocks. To encode
the numerical sequence for each node, a recurrent neural network (RNN) is frequently used due
to its superior performance in predicting time-series data. The literature could be summarized
into two lines of work, long short-term memory (LSTM) based approach and gated recurrent
unit (GRU) based approach.

3.1.1 LSTM Based Approach

As a special form of recurrent neural network, long short-term memory(LSTM) (Hochreiter and
Schmidhuber, 1997) is capable to capture the long-term dependencies and avoid the vanishing
 A Review on GNN in Finance 119

gradient problem. Using memory cells and gate units, it has the following expression:

ft = σ (Wf xt + Uf ht−1 + bf ),
it = σ (Wi xt + Ui ht−1 + bi ),
ot = σ (Wo xt + Uo ht−1 + bo ),
c̃t = tanh(Wc xt + Uc ht−1 + bc ),
ct = fc ◦ ct−1 + it ◦ c̃t−1 ,
ht = ot ◦ tanh(ct ),

where xt ∈ RD is the input vector at time t and D is the number of features, ft , it , ot , c̃t , ct , ht
denotes the forget gate, output gate, cell input, cell state and hidden state vectors respectively,
Wf , Wi , Wo , Wc , Uf , Ui , Uo , Uc are trainable weight matrices and bf , bi , bo , bc are trainable bias
vectors, σ (·)represents the sigmoid activation function, and ◦ denotes the element-wise product.
The hidden state of the LSTM on day t is denoted by: ht = LST M(xt , ht−1 ), s − l  t  s.
Since the LSTM updates the hidden state to capture the structural information, a common
approach to encode the historical data is generating sequential embedding E s using the last
hidden state of LSTM, E s = LSTM(X s ) ∈ Rn×u , where u is the dimension of the output feature.
Then the encoded pricing information is used as input to the graph neural network. For instance,
Chen et al. (2018) used the generated sequential embedding E s as the input feature matrix for
graph convolutional network and Feng et al. (2019) utilized E s as the input feature matrix
for their proposed temporal graph convolutional layer. Using the last hidden state as an input
feature, this type of method could capture the information in the past days while having an
appropriate format to feed into the GNN model.

3.1.2 GRU Based Approach

Gated recurrent unit (GRU) (Cho et al., 2014) is another variant of RNN models. It also applies
gating mechanism and has fewer parameters. It has the following structure:

rt = σ (Wr xt + Ur ht−1 + br ),
zt = σ (Wz xt + Uz ht−1 + bz ),
h̃t = tanh(Wh xt + Uh (rt ◦ ht−1 ) + bh ),
ht = (1 − zt ) ◦ ht−1 + zt ◦ h̃t ,

where xt ∈ RD is the input vector at time t for stock i and D is the number of features, rt ,
zt , h̃t , ht denotes the reset gate, update gate, candidate activation and hidden state vectors
respectively, Wr , Wz , Wh , Ur , Uz , Uh are trainable weight matrices and br , bz , bh are trainable
bias vectors, σ (·) represents the sigmoid activation function, and ◦ denotes the element-wise
product. The hidden states of the GRU on day t is denoted by: ht = GRU(xt , ht−1 ).
Utilizing GRU to encode the past numerical information, we could obtain the hidden state
for each day. Since past days’ impact on the current-day representation may differ, an attention
mechanism is frequently used to assign weights differently. For instance, Sawhney et al. (2020a)
use an additive attention mechanism to aggregate the hidden states across time. Cheng et al.
(2020) utilize the concatenated attention method to incorporate different importance of time.
The attention mechanism aggregates the hidden states of past days and assigns different weights
across time. To get the feature representation at time s, it rewards the influential days when
120 Wang, J. et al.

aggregating the hidden states from time s − l to time s, and thus take temporal dependencies
into account. The obtained node representation is then used as a feature in the graph neural
network model.

3.2 Textual Information

In the financial industry, a large amount of information is of textual form, including financial
news, financial statements, and customer reviews. How to translate the texts into vector repre-
sentation while preserving semantic information, is essential. The following section summarizes
the commonly-used natural language processing (NLP) methodology to convert the unstructured
texts into a vector form.

3.2.1 Word Embedding and Sentence Embedding

Word embedding methods are widely used to represent a word as a fixed-length vector. Then, to
learn the sentence representation, a recurrent neural network model is often utilized to capture
local semantic information. For instance, to embed the news headlines, Li et al. (2020a) encode
the word as word embedding using GloVe (Pennington et al., 2014). Then LSTM method is
applied with an attention mechanism to create the sentence representation. However, since a
word may have multiple meanings, word embedding methods may cause problems by assigning
the same vector to words with different meanings. Thus, there are also approaches to embed at
a sentence level in order to alleviate the problem. For instance, Sawhney et al. (2020a) generate
sentence-level embedding for each Tweet using Universal sentence encoders (Cer et al., 2018).

3.2.2 Language Model

Instead of focusing on generating vectors for words, language models focus on capturing the
pattern of languages to predict the word based on its surrounding words. Taking the contexts
into account, language models achieve the state of art performance on many NLP tasks and
are widely used in the literature. For instance, Liou et al. (2021) use bidirectional encoder
representations from transformers (BERT) model (Devlin et al., 2019) to encode the entire news
article and generated a news embedding. Without modifying the model architecture, the pre-
trained BERT model is able to be fine-tuned and produce state-of-art performance. For example,
FinBERT (Araci, 2019), a BERT model pre-trained specific to the financial domain, is utilized
to encode the text scripts in company earning calls (Sawhney et al., 2020b).

4 Graph Neural Network Models

Proposed by Gori et al. (2005), a graph neural network is a neural network model capable of
processing graphs. Unlike network embedding methods whose major aim is to generate a vector to
represent each node, graph neural network models are designed for a variety of tasks, including
node classification, edge prediction, and graph classification. Due to its wide application and
superior performance, graph neural network models have drawn great attention. In this section,
we present the commonly-used GNN models for each type of graphs, since the methodology may
vary with different graph characteristics. In the supplementary materials, we present a figure
demonstrating the major GNN methodology used for each graph type.
 A Review on GNN in Finance 121

4.1 Homogeneous Graph

Proposed by Kipf and Welling (2017), graph convolutional network (GCN) is a widely-used graph
neural network model and could encode both local graph structure and node features. Extend-
ing the convolution concepts to graphs, graph convolution could be seen as message passing
and information propagation. Aggregating neighbors’ feature information, graph convolutional
networks could be represented with the following layer-wise propagation rule:
1 1
H l+1 = σ (D̃ − 2 ÃD̃ − 2 H l W l ),

where à is the adjacency matrix with added self-connections, D̃ = diag( j Ãij ) W l is trainable
weight matrix of l th layer, H l is the node hidden feature matrix in the l th layer and σ (·) is the
activation function. With its relatively simple model structure and great performance, the GCN
model is often used as a benchmark method to compare with. Among the reviewed literature,
over half of them have applied the GCN method as a benchmark method.
While GCN equally treats the neighbors of the target node, it often occurs that some
neighbors may be more influential than others. Considering various impacts of the neighbor
nodes, Veličković et al. (2018) propose graph attention networks (GAT) and it is able to assign
different weights to nodes in the same neighborhood as follows:

hl+1
i = σ( αijl W l hlj ),
j ∈ Ni

exp(LReLU(a T [W l hli W l hlj ]]))

αijl =  ,
k∈Ni exp(LReLU(a [W hi W hk ]
T l l l l

where hl is the hidden feature vector for node i in the l th layer, W l is trainable weight matrix, a is
a learnable vector, Ni is the neighborhoods of node i, αijl represents attention coefficient of node
j to i at l th layer, σ (·) is the activation function,  denotes vector concatenation, and LReLU
denotes the leaky ReLU activation function. It is also worth mentioning that, since GAT is able
to learn the weights of the neighboring node, we could interpret the learned attention weights
as a relative importance measure, to better understand the model. Similar to GCN, GAT is also
often used as a benchmark method in the reviewed papers with about 40% coverage.

4.2 Directed Graph

A undirected graph has a symmetric adjacency matrix that guarantees a semi-definite Laplacian
matrix, which lays the foundation for applying GCN. The directed graph, on the other hand,
may has asymmetric adjacency matrix and could be handled by spatial-based GNN methods
(Wu et al., 2020), such as GAT. In practice, there is not much work developing methodologies
for directed graph, since it could be processed by spatial-based GNN methods or making the
adjacency matrix symmetric. However, there is also a line of work developing GNN methodology
to predict sequences of events represented as a directed graph.
A sequence of events could be naturally formed as a directed graph, where each node is one
type of event and an edge points from one event to the following one. Given the graph, which
could be seen as a partial sequence of events, how to encode the features and predict the rest of
the sequence is a challenge. The aforementioned GCN and GAT models aim at representation
learning and are used to produce a single output instead of outputting a sequence. To fill the gap,
Li et al. (2016) propose a gated graph neural network (GGNN) that could produce sequential
122 Wang, J. et al.

outputs. It applies gated recurrent unit (GRU) as a recurrent function and is constructed as
follows:

at = AT ht−1 + b,
rt = σ (Wr at + Ur ht−1 ),
zt = σ (Wz at + Uz ht−1 ),
h̃t = tanh(Wh at + Uh (rt ◦ ht−1 )),
ht = (1 − zt ) ◦ ht−1 + zt ◦ h̃t ,

where ht is the updated event representation at t th step, at contains information transferred from
both directions’ edges, rt , zt , h̃t denotes the reset gate, update gate, and candidate activation
vectors at t th respectively, Wr , Wz , Wh , Ur , Uz , Uh are trainable weight matrices, b is a trainable
bias vector, σ (·) is the sigmoid function, ◦ denotes the element-wise product, and tanh denotes
the hyperbolic tangent function. Incorporating the adjacency matrix A, GGNN aggregates the
structural information in every propagation step. Unrolling the recurrence function to a fixed
number, the GGNN ensures convergence without constraining the parameters.

4.3 Bipartite Graph

The aforementioned methods focus on homogeneous graphs whose nodes and edges are all of one
type. In real-life applications, graphs could be heterogeneous. As a running example, in a spam
detection task, we could have a user-item network, where nodes are either users or items. An
edge eij denotes that user i has rated on item j . This type of graph is well known as a bipartite
graph G with the following notation: G = (V, E), where V = {U, W } is a set of nodes and could
be divided into two non-empty and disjoint sets U, W. E is a set of edges, where every edge joins
a node in U and a node in W.
Based on the characteristics of a bipartite graph, the commonly-used methodologies could
be categorized into a framework as follows. In each iteration, the edge information hluw is updated
aggregating its previous hidden state hl−1 l−1 l−1
uw and hidden states of the two nodes it links (hu , hw ),
as shown in equation (1). At each iteration, a user node u aggregates information from its rated
items hlN (u) and its past hidden state hl−1
u , as in equation (3). While the representation of the
l
rated items hN (u) is updated aggregating the information from the linking edges hluw and the
item nodes hl−1w , as in equation (2). Then, the item node w is updated respectively as shown in
the following equations:

hluw = σ (WEl · AGGE (hl−1 l−1 l−1

uw , hu , hw )), (1)
hN (u) =
l
σ (WN (U ) · AGGU (AGGU W (hl−1
l l
w , huw ))), ∀w ∈ {N (u)}, (2)
hlu = concat(WUl · hl−1 l
u , hN (u) ), (3)
hlN (w) = l
σ (WN (W ) · AGGW (AGGW U (hu , huw ))),
l−1 l
∀u ∈ {N (w)}, (4)
hlw = concat(WWl−1 · hlw , hlN (w) ), (5)

where euw represents the edge representation for edge linking node u and w, hluw , hlu , hlw , hlN (u) ,
hlN (w) are hidden states at l th layer, WEl , WUl , WN
l l l
(U ) , WN (W ) , WW are trainable weight matrices
at l th layer, AGGE (·), AGGU (·), AGGU W (·), AGGW (·), AGGW U (·) are user-chosen aggregation
functions, σ (·) is the activation function, and concat denotes the vector concatenation.
 A Review on GNN in Finance 123

There exist many modeling methodologies for bipartite graphs that could fit into the above
framework. For instance, Zhang et al. (2020) propose a similar model structure as the frame-
work and apply the attention mechanism as the aggregation function, since different items may
have different impacts when learning uses’ representations. Instead of using all neighbors, Li
et al. (2019a) utilize a sampling technique when aggregating neighbors’ information in each it-
eration. There are also literature using the above framework as a building block and combining
clustering methodology to learn a hierarchical representation of the graph, since hierarchical
representation with various GNN models could achieve satisfactory performance. For instance,
Li et al. (2019b) utilize the node embedding generated from the framework to cluster users into
different communities and make a recommendation based on both community information and
user information. Specifically, the user information is decomposed into two orthogonal spaces
representing community-level information and individualized user preferences. Li et al. (2020c)
treat the framework as a GNN module and stack it in a hierarchical fashion. With the embed-
ding generated from the framework, clustering algorithms are performed to generate a coarsened
graph which is used as an input for the next GNN layer.

4.4 Multi-Relation Graph

Instead of a simple homogeneous graph where all nodes and edges have the same type, in real life,
there may exist multiple relations between nodes. For example, in a malicious account detection
task, we could construct an Amazon review network. Users are the nodes in the graph and there
are three relations between users: reviewing the same product, having the same star rating, and
having similar texts. The multi-relation graph is denoted as G: G = (V, E1:R ), where V is the set
of nodes, E1:R is the set of edges, R is the number of node types, ei,jr
∈ Er is an edge between
node i, j with a relation r ∈ {1, . . . , R}.
A frequently used approach is to transform the heterogeneous graph into multiple homoge-
neous graphs by extracting R subgraphs {G r = (V, Er ), r = 1, . . . , R} from it. Each subgraph
G r only preserves one type of edge and thus is homogeneous. This line of work could be unified in
a two-step framework. The first step implements the sub-graph aggregation to aggregate neigh-
bor information in each sub-graph as shown in equation (6). The second step is to conduct the
inter-relation aggregation to aggregate relation-specific embeddings as equation (7),

hli,r = f (AGGr {hl−1

j,r }), ∀j s.t.(i, j ) ∈ Er , (6)
hli = i }),
g(AGG{hli,(1:R) , hl−1 (7)

where hli,r is the subgraph-specific embedding of node i in subgraph r in l th layer, hli is the general
embedding of node i in l th layer, AGGr (·) is the aggregation function in subgraph r, AGG(·) is
the inter-relation aggregation function, and f (·), g(·) are user-defined functions.
There are multiple methods that could be categorized into the above two-step framework.
For instance, in a fraud classification task, Liu et al. (2018) observe that fraudsters tend to
congregate in topology and thus use weighted sum for within-relation aggregation to capture
this congregation pattern. They then apply an attention mechanism for inter-relation aggregation
to learn the significance for each sub-graph as follows:

hli,r = σ (Weighted mean{hl−1

j,r }), ∀j s.t.(i, j ) ∈ Er ,
hli = σ (Xi W + Attention{hli,(1:R) }),
124 Wang, J. et al.

where Xi is the feature vector for node i, W is a trainable matrix, σ (·)is the activation function,
and Attention denotes the attention aggregator.
To incorporate neighbors’ information, Dou et al. (2020) use the mean aggregator for within-
relation aggregation. To reduce the computational cost and keep the relational importance infor-
mation, they apply a pre-calculated parameter prl as the weight in the intra-relation aggregation
step. They also test several aggregating functions when aggregating relation-specific embeddings
with the following structure:

hli,r = σ (Mean{hl−1
j,r }), ∀j s.t.(i, j ) ∈ Er ,
hli = σ (hl−1
i + AGG{hli,r · prl }), ∀r ∈ (1, . . . , R),

where prl is a pre-trained weight.

Since different relations provide various facets of user characteristics, relationship-specific
embedding may have different statistical properties, which may cause trouble when aggregating
them in a lower-level space. To deal with that, Wang et al. (2019) project the relation-specific
node embedding to higher spaces using multi-layer perception (MLP) and concatenate them
with relation-level attention:

hli,r = MLP{hl−1
i,r }, where hi,r = Attention{xj,r : ∀j s.t.(i, j ) ∈ Er },
1

hli = Concatenation with attention{hli,(1:R) },

where Attention denotes the attention aggregator.

4.5 Dynamic Graph

The previously mentioned neural network models generally focus on a static graph. However,
in real-life settings, a graph may be dynamically evolving since relations may be updated with
time. For example, in order to predict the loan default risk, a guarantee network needs to be
updated, adding newly-constructed guarantee relationships and removing companies that have
fully paid the loan. With the rapid development of graph neural network methodologies on
static graphs, there emerges a trend to extend GNN models to a dynamic setting. Consider a
sequence of T graphs G seq = {G1 , . . . , GT }, where Gi = (Vi , Ei ) represent the graph at ith time
point. The feature matrices are represented as X = {X1 , . . . , XT } and the adjacency matrices
are A = {A1 , . . . , AT }.
To capture the sequential pattern in the dynamic graph, a common approach is to train a
GNN to generate the node embedding at each time stamp and then utilize a recurrent neural
network to aggregate the information. For example, Cheng and Zhang L Wang X (2020b) obtain
the node embeddings at each time step by training a GCN with multi-head attention. Then, they
utilize the GRU to capture the sequential pattern with a temporal attention layer to capture
the temporal variation over timestamps. Similarly, Yang et al. (2020) first aggregate node and
edge information in each snapshot and then employ a LSTM operator to capture the temporal
variations in the node embeddings.
In the aforementioned methods, a graph neural network is learned for feature aggregation
and an RNN model is trained to capture the sequential pattern of the node embeddings. However,
in reality, a node may appear and disappear, which may worsen the performance of the RNN
model when updating the node representation. In a guarantee network, for example, a company
that has borrowed loans could disappear from the graph after it pays all the debts and could
 A Review on GNN in Finance 125

appear again backing up other companies’ loans. To overcome the limitation, Pareja et al. (2020)
proposed EvolveGCN which utilizes a recurrent neural network to evolve the GCN parameters
instead of updating the node embeddings. For each time point t, a GCN model is constructed
as follows to fit the graph Gt :
−1 −1
Htl+1 = σ (D̃t 2 Ãt D̃t 2 Htl Wtl ),

where Ãt is the adjacency matrix with added self-connections, D̃t = diag( j Ãij ) is the degree
matrix, Wtl is trainable weight matrix of l th layer, Htl is the matrix of activation in the l th layer,
and σ (·) is the activation function.
To update the weight matrix Wtl , Pareja et al. (2020) propose two methods. The first method
considers Wtl as a hidden state of the dynamics and update it using a GRU model, as shown in
equation (8). The second method treats Wtl as an output state which is updated using a LSTM
method, as shown in equation (9). The structure for both methods is as follows:

Wtl = GRU(Htl , Wt−1

l
), (8)
Wtl = LSTM(Wt−1
l
). (9)

Compared to the second method, the first method incorporates the updated node embedding
in the recurrent neural network and it may lead to better performance when node features are
informative.

5 Application
In this section, we have detailed some financial applications that the GNN methods have been
commonly applied on. We have also summarized features, graphs, methods, evaluation metrics,
and baselines used in each financial application in the supplementary materials.

5.1 Stock Movement Prediction

Though there are still debates on whether stocks are predictable, stock prediction receives great
attention and there are rich literature on predicting stock movements using machine learning
methods. However, the task of stock prediction is challenging due to the volatile and non-linear
nature of the stock market. Traditionally, there are two major approaches to handle the task:
technical analysis and fundamental analysis (Sawhney et al., 2020a). Technical analysis utilizes
numerical features such as closing prices and trading volumes, while the fundamental analysis
approach includes non-numerical information, such as news and earning calls. The limitation
of these non-graph approaches is that they often have a hidden assumption that the stocks are
independent. To take the dependence into account, there is an increasing trend to represent the
stock relations in a graph where each stock is represented as a node and an edge would exist
if there are relations between two stocks. Predicting multiple stocks’ movements could then be
formed as a node classification task and GNN models could be utilized to make the prediction.
However, there also exist challenges to apply the GNN methods in the stock prediction task.
Unlike other fields where the benchmark graphs are available, to the best of our knowledge, there
is no off-the-shelf graph representing inter-stock relations. With abundant relations existing in
the financial system, it becomes challenging to obtain and select the relation for graph construc-
tion. Moreover, owing to the volatility of the stock market, how to model the sequential features
126 Wang, J. et al.

and capture the temporal patterns are also critical. Also, the financial industry has rich data
sources including financial statements, news and pricing information, which impose difficulty on
modeling the data.
There are multiple ways to construct the stock relational graph. For instance, believing that
correlation on historical prices reflects the inter-stock relation, Li et al. (2020a) construct the
graph using the correlation matrix of historic data to predict the movement of Tokyo stock price
index. On the other hand, Matsunaga et al. (2019) borrow information from knowledge bases
and construct supplier, customer, partner, and shareholder relational graphs. With multiple ways
of graph construction, there doesn’t exist a “best” graph due to the lack of graph evaluation
methods. Future work could be done to design a graph evaluation method to help researchers
better construct a relational graph.
To effectively process the sequential data and incorporate related corporations’ information
Chen et al. (2018) propose a joint model using LSTM and GCN to predict the stock movement.
However, Chen et al. (2018)’s approach assumes that the relations between stocks are static,
which may not reflect the reality. Instead, Feng et al. (2019) propose a temporal graph convo-
lution layer to capture the stock relations in a time-sensitive manner, so that the strength of
relation could be evolving over time. The relations are then updated based on historical pricing
sequences and the proposed method obtained better performance compared to GCN. Believing
that stock description documents also contain information reflecting the changes in companies’
effect, Ying et al. (2020) capture the temporal relation by both sequential features and stock
document attributes with a time-aware relational attention network.
The aforementioned methods focus on capturing the temporal dependencies, while Sawhney
et al. (2020a) focus on fusing data from different sources. Sawhney et al. (2020a) propose a
multipronged attention network to jointly learn from historical price, social media, and inter
stock relations. Encoded pricing and textual information are used as node feature inputs to
GAT, where the graph information comes from the Wiki company-based relations. The attention
mechanisms is applied to allocate different weights on various data sources and latent correlations
may learned via the attention layers.

5.2 Loan Default Risk Prediction

For commercial banks and financial regularity institutions, monitoring and assessing the default
risk is at the heart of risk controlling process. As one of the credit risks, default risk is the
probability that the borrower fails to pay the interest and principal on time. With a binary
outcome, loan default prediction could be seen as a classification problem and is commonly
addressed utilizing user-related features with classifiers including neural network (Turiel and
Aste, 2020) and gradient boosted trees (Ma et al., 2018). Since the probability that a borrower
defaults may be influenced by other related individuals, there is plenty of literature forming a
graph to reflect the interactions between borrowers. With the rapid growth of GNN methods,
GNN methods are widely applied on the graph structure for loan default predicting problems.
There are currently three lines of work focusing on various types of loans: guarantee loans,
e-commerce loans, and other loans.
The guarantee loan allows small entrepreneurs to back each other in order to increase their
credibility. It has a debt obligation contract that specifies that if one corporation fails to pay
the debt, its guarantor needs to pay for it. A guarantee network naturally arises where each
node is a company and directed edges represent the guarantee relationship. To learn a better
representation of the network, Cheng et al. (2019) utilize the graph attention layer and design
 A Review on GNN in Finance 127

a objective function, so that vertices with similar structures will be closer in the learned feature
space. Since the guarantee relations changes with time, Cheng and Zhang L Wang X (2020b)
forms a dynamic guarantee network to represent the dynamics. A recurrent graph neural network
layer is developed to learn the temporal pattern and attentional weights are learned for each
time point via an attention architecture.
Unlike guarantee loans that the loan information could be naturally represented in a di-
rected graph, other loan types may not have a clear graph structure and researchers need to
construct the graph based on interactive information. For instance, Xu et al. (2021) construct
a user relation graph where users are connected by various relationships, such as social con-
nections, transactions, and device usage. However, the interactive graph may also contain noisy
data, which may be irrelevant. Since the massive interactive information may be noisy and
the impacting supply-chain information is deficient, Yang et al. (2020) extract supply-chain
relations while predicting loan defaults. Forming the interaction data as a graph, Yang et al.
(2020) formulate the supply chain mining task as a link prediction task and thus construct a
supply chain network, which is then used to predict the default probability with GNN method-
ology.
With the rise of e-commerce, e-commerce consumer lending service is gaining popularity to
enhance consumers’ purchasing power. Able to obtain information from multiple facets, the e-
commerce platform could have multi-view data and multi-relation networks, which may require
sophisticated modeling methodology. For instance, in order to predict the default probability
for each consumer with multi-view data, Liang et al. (2021) utilize a hierarchical attention
mechanism to encode the features on each view. Exploring multiplex relations, Hu et al. (2019)
propose an attributed multiplex graph-based model with relation-specific layer and attention
mechanism to jointly model multiple relations. To simultaneously model the labeled and unla-
beled data, Wang et al. (2019) proposed a semi-supervised graph neural network approach and
obtain interpretable results.

5.3 Recommender System of e-Commerce

With the rapid growth of e-commerce, customers gradually get used to shopping online and
are exposed to a numerous range of products. To alleviate the burden of users choosing the
appropriate item, the recommender system is developed to suggest products to users based on
predicted item ratings. Presenting the user and item information in a graph, GNN methodologies
are widely used in recommender system related tasks including click rate prediction and fake
review detection.
To accurately predict users’ preferences and recommend the appropriate item, it is vital
to exploit information for users, items, and their interactions. A widely-used representation
of that information is a bipartite graph, where nodes are of two types, user and item, and
edges represent that there are relationships between user and item nodes. Noticing that the
community the user belongs to may affect the shopping decision, for example, a user belongs
to the traveler group may purchase travel-related items, Li et al. (2019b) combine the bipartite
graph modeling algorithm with clustering techniques to reflect the community impact and the
individual preference. However, Li et al. (2019b)’s approach only considers the hierarchy in the
user side, while items may also have hierarchical information. To fill the gap, Li et al. (2020c)
stack several GNN modules hierarchically to capture the hierarchical structure in both user and
item perspectives. Embeddings learned from the GNN layer are clustered and used as an input
for the next GNN layer, which could preserve the high-order hierarchical connections.
128 Wang, J. et al.

The recommender system is mainly based on the past history of the user, including its rating
and reviews on the item. However, fake ratings and feedback may be posted by the fraudsters
to seek financial benefits. To detect fraudulent reviews on the e-commerce platform, Kudo et al.
(2020) construct a directed and signed comment graph with a signed graph convolutional network
approach. Compared to Kudo et al. (2020)’s approach which only considers the comment graph,
Li et al. (2019a) integrate both the bipartite user-item graph and a comment graph to capture the
local and global context of the comments. Noticing that camouflage behaviors of the fraudsters
may deteriorate the performance of fraud detection mechanism and has seldom been considered
by prior works, Dou et al. (2020) propose a model against both feature and relation camouflage.
For each node, only informative neighbors are selected for the next aggregation step, utilizing
a similarity measure and a reinforcement learning mechanism. While the above literature focus
on the fraud review detection side, there are also works that accomplish both fraud review
detection and item recommendation tasks. For example, Zhang et al. (2020) propose a GCN-
based framework that performs both item recommendation and fraud detection in an end-to-end
manner, while each of the tasks is beneficial for the other one.

5.4 Fraud Detection

Including payment fraud, identity theft, financial scam and insurance fraud, financial fraud has a
variety of types and has an increasing trend (Kurshan and Shen, 2020). Observing that fraudsters
tend to have abnormal connectivity with other users, there is a trend to present users’ relations
in a graph and thus, the fraud detection task could be formulated as a node classification task.
Aiming to detect the malicious accounts, who may attack the online services to seek excessive
profits, Liu et al. (2018) find out that fraudsters have two patterns: device aggregation and
activity aggregation. Due to economic constraints, attackers tend to use limited number of
devices and perform activities in a limited time, which may be reflected in the local graph
structure. With this observation, Liu et al. (2018) propose a variant of GCN and use sum
operators to capture the aggregation pattern. While the malicious accounts may be aggregated
together, Liu et al. (2019) argue that the normal accounts could also be connected with the
malicious account and may be mislabeled as malicious, which adds noisy signals to the graph.
Taking into account that the graph could be noisy and nodes may have different impact, Liu
et al. (2019) propose an adaptive path layer to adaptively select the important neighbor nodes
that contribute most to the target node. Applying the GNN methodologies proposed by Liu
et al. (2018), Liang et al. (2019) stack multiple adaptive path layers to aggregate neighbors’
features and have great performance in a insurance fraud detection task.
The above literature focus on a bipartite graph, where nodes are either account or devices,
and there are also literature utilizing other types of graphs. For instance, Jiang et al. (2019)
construct a homogeneous user network connecting user nodes based on their similarity of be-
haviors and apply the GCN model. Rao et al. (2020) focus on a dynamic graph of registration
records and implement GCN layers on structural and temporal subgraphs.
Besides the literature developing GCN-based methodologies, there are also fraud-detection
related literature applying GNN models with other structures. Since GCN may suffer from over-
smoothing problem and have shallow model structure, Lv et al. (2019) propose to replace the
graph convolutional matrix with auto-encoder to increase the depth of the neural network. Zhao
et al. (2021) argues that GNN models with random-walk based losses have poor performance in
anomaly detection task, since nodes of the same label may not be closer. They then propose a
loss function which leads to better performance and has bounded prediction error.
 A Review on GNN in Finance 129

5.5 Event Prediction

Financial events, including revenue growth, acquisition and bankruptcy, could provide valuable
information on market trends and could be used to predict future stock movement. Therefore, it
draws great attention on how to predict next financial event based on past events and currently
GGNN model is often used to accomplish the task. For example, given a sequence of financial
events of Chinese listed companies, Yang et al. (2019) aim to predict the next event type and
construct an event graph where each node is a financial event and edges are weighted using the
frequency of the event pairs. Harl et al. (2020) transform a binary classification problem into an
event prediction task by dividing the loan application process into several events and predicting
whether the next event would be accepting or rejecting the application. Utilizing the GGNN
model, they obtain the predictions with high accuracy.

6 Challenges
6.1 Graph Evaluation Methods
To justify the inclusion of a graph, a commonly-used evaluation method is to compare the
outcome for a graph-based machine learning method with a graph-free machine learning method
(Feng et al., 2019; Li et al., 2020a). However, there is little discussion on comparing different
graphs’ effects and quality, while existing literature discussing graph comparisons are often
inadequate. For example, Liang et al. (2019) visualize the structural patterns in different graphs,
concluding that the device sharing graph is more appropriate based on the observed patterns.
Without presenting the evaluation metric for each graph, the graph comparison based on the
visualization may not be adequate. The problem may be more severe in similarity-based graph
construction since a threshold needs to be set to determine whether an edge exists. Different
threshold values may lead to completely different graphs and thus affect the model performance.
Thus, justification on threshold setup during graph construction is of great importance. Some
efforts have been made to tackle this problem, such as utilizing a reinforcement learning approach
to automatically select the optimal threshold (Dou et al., 2020). More attention may need to be
drawn to develop a framework to assess the graph quality systematically.

6.2 Explainability
Combining both graph structural information and feature information, GNN models are often
complicated and it is challenging to make an interpretation. Recently, there are some literature
focusing on the explainability of GNN models. GNNExplainer (Ying et al., 2019), for example, is
proposed to provide an interpretable explanation on trained GNN models such as GCN and GAT.
Model explainability in financial tasks is of great importance, since understanding the model
could benefit decision-making and reduce economic losses. However, there is little literature
studying the explainability of GNN models in a financial application, which often accompanies
by heterogeneous and dynamic graphs. Current literature focuses on relatively simpler graphs.
For example, Li et al. (2019c) extend the GNNExplainer to a weighted directed graph and apply
it on a Bitcoin transaction graph. Rao et al. (2020) propose an explainable fraud prediction
system that could operate on heterogeneous graphs consisting of different node and edge types.
More work could be done on the explainability of GNN models with edge-attributed graphs and
dynamic graphs, which are not yet considered.
130 Wang, J. et al.

6.3 Task Type

Tasks of GNN models are commonly classified into three categories: node-level task, edge-level
task, and graph-level task (Wu et al., 2020). However, the GNN methods applied on financial
applications are mostly focused on the node-level task. For example, the stock movement predic-
tion task is often formulated as a node classification task, where stocks are represented as nodes.
There exist some literature focus on other types of task, for instance, Yang et al. (2020) aim to
mine the underlying supply chain network and formulate it as a link prediction task, but this line
of work is rare. Since there are rich literature developing GNN methodologies on edge prediction
or graph classification tasks, there are rich opportunities on applying recently-developed GNN
methods on financial fields if financial tasks could be formulated into an edge or graph level
task.

6.4 Data Availability

To pursue reproducibility, it is a common practice to release both data sources and codes when
publishing the paper. However, in the reviewed papers, only about 24% of them release the
code. Since the financial tasks are commonly based on real-world problems, data may have some
restrictions due to the privacy obligation of the related corporations. Lack of open-source codes
and datasets, it is hard to reproduce previous works and compare their methodologies in the
latter literature. Thus, it is of great value to construct benchmark datasets, so that methods
could be compared on the same data.

6.5 Scalability
In the real-world financial scenario, commercial data are often of large scales. For instance,
Yang (2019) utilize data from a popular e-commerce platform and it contains about 483 million
nodes with 231 million edges. How to improve the scalability of GNNs is vital but challenging.
Computing the Laplacian matrix becomes hard with millions of nodes and for a graph of irregular
Euclidean space, optimizing the algorithm is also difficult. Sampling techniques may partially
solve the problem with the cost of losing structural information. Thus, how to maintain the
graph structure and improve the efficiency of GNN algorithms are worth further exploration.

Supplementary Material
In the supplementary materials, we present materials that are not covered in the main text.
The supplementary materials contain the summary table for each financial application, figures
categorizing major GNN methodologies for each graph type and acronyms used in the text.

References
Araci D (2019). Finbert: Financial sentiment analysis with pre-trained language models. arXiv
preprint: https: // arxiv. org/ abs/ 1908. 10063 .
Cer D, Yang Y, Kong S, Hua N, Limtiaco N, John RS et al. (2018). Universal sentence encoder
for english. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing: System Demonstrations, 169–174.
 A Review on GNN in Finance 131

Chen Y, Wei Z, Huang X (2018). Incorporating corporation relationship via graph convolutional
neural networks for stock price prediction. In: Proceedings of the 27th ACM International
Conference on Information and Knowledge Management, 1655–1658.
Cheng D, Niu Z, Zhang Y (2020). Contagious chain risk rating for networked-guarantee loans.
In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery
& Data Mining, 2715–2723.
Cheng D, Tu Y, Ma Z, Niu Z, Zhang L (2019). Risk assessment for networked-guarantee loans
using high-order graph attention representation. In: Proceedings of the Twenty-Eighth Inter-
national Joint Conference on Artificial Intelligence, IJCAI-19, 5822–5828. International Joint
Conferences on Artificial Intelligence Organization.
Cheng D, Wang X, Zhang Y, Zhang L (2020). Risk guarantee prediction in networked-loans.
In: Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence,
IJCAI-20 (C Bessiere, ed.), 4483–4489. International Joint Conferences on Artificial Intelli-
gence Organization. Special Track on AI in FinTech.
Cho K, van Merrienboer B, Gulcehre C, Bougares F, Schwenk H, Bengio Y (2014). Learning
phrase representations using rnn encoder-decoder for statistical machine translation. In: Con-
ference on Empirical Methods in Natural Language Processing (EMNLP 2014).
Devlin J, Chang MW, Lee K, Toutanova K (2019). BERT: Pre-training of deep bidirectional
transformers for language understanding. In: Proceedings of the 2019 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Tech-
nologies, Volume 1 (Long and Short Papers), 4171–4186. Association for Computational Lin-
guistics, Minneapolis, Minnesota.
Dou Y, Liu Z, Sun L, Deng Y, Peng H, Yu PS (2020). Enhancing graph neural network-based
fraud detectors against camouflaged fraudsters. In: Proceedings of the 29th ACM International
Conference on Information & Knowledge Management, 315–324.
Feng F, He X, Wang X, Luo C, Liu Y, Chua TS (2019). Temporal relational ranking for stock
prediction. ACM Transactions on Information Systems (TOIS), 37(2): 1–30.
Gori M, Monfardini G, Scarselli F (2005). A new model for learning in graph domains. In: Pro-
ceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005, volume 2,
729–734. IEEE.
Harl M, Weinzierl S, Stierle M, Matzner M (2020). Explainable predictive business process
monitoring using gated graph neural networks. Journal of Decision Systems, 1–16.
Hochreiter S, Schmidhuber J (1997). Long short-term memory. Neural computation, 9(8):
1735–1780.
Hu B, Zhang Z, Shi C, Zhou J, Li X, Qi Y (2019). Cash-out user detection based on attributed
heterogeneous information network with a hierarchical attention mechanism. In: Proceedings
of the AAAI Conference on Artificial Intelligence, volume 33, 946–953.
Hu B, Zhang Z, Zhou J, Fang J, Jia Q, Fang Y, et al. (2020). Loan default analysis with multiplex
graph learning. In: Proceedings of the 29th ACM International Conference on Information &
Knowledge Management, 2525–2532.
Huang J, Chai J, Cho S (2020). Deep learning in finance and banking: A literature review and
classification. Frontiers of Business Research in China, 14: 1–24.
Jiang J, Chen J, Gu T, Choo KKR, Liu C, Yu M, et al. (2019). Anomaly detection with graph
convolutional networks for insider threat and fraud detection. In: MILCOM 2019-2019 IEEE
Military Communications Conference (MILCOM), 109–114. IEEE.
Jiang W (2021). Applications of deep learning in stock market prediction: Recent progress.
132 Wang, J. et al.

Expert Systems with Applications, 115537.

Kipf TN, Welling M (2017). Semi-supervised classification with graph convolutional networks.
In: International Conference on Learning Representations (ICLR).
Kudo W, Nishiguchi M, Toriumi F (2020). Gcnext: Graph convolutional network with expanded
balance theory for fraudulent user detection. Social Network Analysis and Mining, 10(1): 1–12.
Kurshan E, Shen H (2020). Graph computing for financial crime and fraud detection: Trends,
challenges and outlook. International Journal of Semantic Computing, 14(04): 565–589.
Lee JB, Rossi RA, Kim S, Ahmed NK, Koh E (2019). Attention models in graphs: A survey.
ACM Transactions on Knowledge Discovery from Data (TKDD), 13(6): 1–25.
Lee JW, Lee WK, Sohn SY (2021). Graph convolutional network-based credit default prediction
utilizing three types of virtual distances among borrowers. Expert Systems with Applications,
168: 114411.
Li A, Qin Z, Liu R, Yang Y, Li D (2019a). Spam review detection with graph convolutional
networks. In: Proceedings of the 28th ACM International Conference on Information and
Knowledge Management, 2703–2711.
Li C, Jia K, Shen D, Shi CJR, Yang H (2019b). Hierarchical representation learning for bipartite
graphs. In: IJCAI, 2873–2879.
Li W, Bao R, Harimoto K, Chen D, Xu J, Su Q (2020a). Modeling the stock relation with graph
network for overnight stock movement prediction. In: no. CONF, 4541–4547.
Li X, Liu S, Li Z, Han X, Shi C, Hooi B, et al. (2020b). Flowscope: Spotting money laundering
based on graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence, volume
34, 4731–4738.
Li X, Saúde J, Reddy P, Veloso M (2019c). Classifying and understanding financial data using
graph neural network. In: AAAI Workshop on Knowledge Discovery from Unstructured Data
in Financial Services 2020.
Li Y, Zemel R, Brockschmidt M, Tarlow D (2016). Gated graph sequence neural networks.
In: Proceedings of ICLR’16.
Li Z, Shen X, Jiao Y, Pan X, Zou P, Meng X, et al. (2020c). Hierarchical bipartite graph neural
networks: Towards large-scale e-commerce applications. In: 2020 IEEE 36th International
Conference on Data Engineering (ICDE), 1677–1688. IEEE.
Liang C, Liu Z, Liu B, Zhou J, Li X, Yang S, et al. (2019). Uncovering insurance fraud conspiracy
with network learning. In: Proceedings of the 42nd International ACM SIGIR Conference on
Research and Development in Information Retrieval, 1181–1184.
Liang T, Zeng G, Zhong Q, Chi J, Feng J, Ao X, et al. (2021). Credit risk and limits forecast-
ing in e-commerce consumer lending service via multi-view-aware mixture-of-experts nets.
In: Proceedings of the 14th ACM International Conference on Web Search and Data Mining,
229–237.
Liou YT, Chen CC, Tang TH, Huang HH, Chen HH (2021). Finsense: An assistant system for
financial journalists and investors. In: Proceedings of the 14th International Conference on
Web Search and Data Mining.
Liu Z, Chen C, Li L, Zhou J, Li X, Song L, et al. (2019). Geniepath: Graph neural networks with
adaptive receptive paths. In: Proceedings of the AAAI Conference on Artificial Intelligence,
volume 33, 4424–4431.
Liu Z, Chen C, Yang X, Zhou J, Li X, Song L (2018). Heterogeneous graph neural networks
for malicious account detection. In: Proceedings of the 27th ACM International Conference on
Information and Knowledge Management, 2077–2085.
 A Review on GNN in Finance 133

Lv L, Cheng J, Peng N, Fan M, Zhao D, Zhang J (2019). Auto-encoder based graph convolu-
tional networks for online financial anti-fraud. In: 2019 IEEE Conference on Computational
Intelligence for Financial Engineering & Economics (CIFEr), 1–6. IEEE.
Ma X, Sha J, Wang D, Yu Y, Yang Q, Niu X (2018). Study on a prediction of p2p network loan
default based on the machine learning lightgbm and xgboost algorithms according to different
high dimensional data cleaning. Electronic Commerce Research and Applications, 31: 24–39.
Matsunaga D, Suzumura T, Takahashi T (2019). Exploring graph neural networks for stock
market predictions with rolling window analysis. arXiv preprint: https: // arxiv. org/ abs/
1909. 10660 .
Ozbayoglu AM, Gudelek MU, Sezer OB (2020). Deep learning for financial applications: A
survey. Applied Soft Computing, 106384.
Pareja A, Domeniconi G, Chen J, Ma T, Suzumura T, Kanezashi H, et al. (2020). Evolvegcn:
Evolving graph convolutional networks for dynamic graphs. In: Proceedings of the AAAI Con-
ference on Artificial Intelligence, volume 34, 5363–5370.
Pennington J, Socher R, Manning CD (2014). Glove: Global vectors for word representation.
In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing
(EMNLP), 1532–1543.
Rao SX, Zhang S, Han Z, Zhang Z, Min W, Cheng M, Shan Y, Zhao Y, Zhang C (2020).
Suspicious massive registration detection via dynamic heterogeneous graph neural networks.
arXiv preprint: https: // arxiv. org/ abs/ 2012. 10831 .
Rao SX, Zhang S, Han Z, Zhang Z, Min W, Chen Z, et al. (2020). xfraud: Explainable fraud
transaction detection on heterogeneous graphs. arXiv preprint: https: // arxiv. org/ abs/
2011. 12193 .
Sawhney R, Agarwal S, Wadhwa A, Shah R (2020a). Deep attentive learning for stock move-
ment prediction from social media text and company correlations. In: Proceedings of the 2020
Conference on Empirical Methods in Natural Language Processing (EMNLP), 8415–8426.
Sawhney R, Khanna P, Aggarwal A, Jain T, Mathur P, Shah R (2020b). Voltage: Volatility fore-
casting via text-audio fusion with graph convolution networks for earnings calls. In: Proceed-
ings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP),
8001–8013.
Turiel J, Aste T (2020). Peer-to-peer loan acceptance and default prediction with artificial
intelligence. Royal Society Open Science, 7(6): 191649.
Veličković P, Cucurull G, Casanova A, Romero A, Liò P, Bengio Y (2018). Graph attention
networks. In: International Conference on Learning Representations.
Vrandečić D, Krötzsch M (2014). Wikidata: A free collaborative knowledgebase. Communica-
tions of the ACM, 57(10): 78–85.
Wang D, Lin J, Cui P, Jia Q, Wang Z, Fang Y, et al. (2019). A semi-supervised graph attentive
network for financial fraud detection. In: 2019 IEEE International Conference on Data Mining
(ICDM), 598–607. IEEE Computer Society.
Weber M, Domeniconi G, Chen J, Weidele DKI, Bellei C, Robinson T, et al. (2019). Anti-money
laundering in bitcoin: Experimenting with graph convolutional networks for financial forensics.
arXiv preprint: https: // arxiv. org/ abs/ 1908. 02591 .
Wu Z, Pan S, Chen F, Long G, Zhang C, Philip SY (2020). A comprehensive survey on graph
neural networks. In: IEEE Transactions on Neural Networks and Learning Systems, 4–24.
IEEE.
Xu B, Shen H, Sun B, An R, Cao Q, Cheng X (2021). Towards consumer loan fraud detection:
134 Wang, J. et al.

Graph neural networks with role-constrained conditional random field. In: Proceedings of the
AAAI Conference on Artificial Intelligence, volume 35, 4537–4545.
Xu K, Hu W, Leskovec J, Jegelka S (2018). How powerful are graph neural networks? arXiv
preprint: https: // arxiv. org/ abs/ 1810. 00826 .
Yang H (2019). Aligraph: A comprehensive graph neural network platform. In: Proceedings of
the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,
3165–3166.
Yang S, Zhang Z, Zhou J, Wang Y, Sun W, Zhong X, et al. (2020). Financial risk analysis for
smes with graph-based supply chain mining. In: Proceedings of the Twenty-Ninth International
Joint Conference on Artificial Intelligence, 4661–4667.
Yang Y, Wei Z, Chen Q, Wu L (2019). Using external knowledge for financial event prediction
based on graph neural networks. In: Proceedings of the 28th ACM International Conference
on Information and Knowledge Management, 2161–2164.
Ying R, Bourgeois D, You J, Zitnik M, Leskovec J (2019). GNN explainer: A tool for post-hoc
explanation of graph neural networks. CoRR, arXiv preprint: https://arxiv.org/abs/abs/
1903.03894.
Ying X, Xu C, Gao J, Wang J, Li Z (2020). Time-aware graph relational attention network
for stock recommendation. In: Proceedings of the 29th ACM International Conference on
Information & Knowledge Management, 2281–2284.
Zhang M, Chen Y (2018). Link prediction based on graph neural networks. Advances in Neural
Information Processing Systems, 31.
Zhang S, Tong H, Xu J, Maciejewski R (2019). Graph convolutional networks: a comprehensive
review. Computational Social Networks, 6(1): 1–23.
Zhang S, Yin H, Chen T, Hung QVN, Huang Z, Cui L (2020). Gcn-based user representation
learning for unifying robust recommendation and fraudster detection. In: Proceedings of the
43rd International ACM SIGIR Conference on Research and Development in Information
Retrieval, 689–698.
Zhao T, Deng C, Yu K, Jiang T, Wang D, Jiang M (2021). Gnn-based graph anomaly detection
with graph anomaly loss. In: The Second International Workshop on Deep Learning on Graphs:
Methods and Applications (DLG-KDD’20). Available at: https://deep-learning-graphs.
bitbucket.io/dlg-kdd20/.
Zhou J, Cui G, Hu S, Zhang Z, Yang C, Liu Z, et al. (2020). Graph neural networks: A review
of methods and applications. AI Open, 1: 57–81.
Zhu YN, Luo X, Li YF, Bu B, Zhou K, Zhang W, et al. (2020). Heterogeneous mini-graph
neural network and its application to fraud invitation detection. In: 2020 IEEE International
Conference on Data Mining (ICDM), 891–899. IEEE.