On The Combination of Graph Data For Assessing Thin-File Borrowers' Creditworthiness

On the combination of graph data for assessing thin-file
borrowers’ creditworthiness
Ricardo Muñoz-Cancino1 , Cristián Bravo2 , Sebastián A. Rı́os3 , and Manuel Graña4
1,3 Business
Intelligence Research Center (CEINE), Industrial Engineering Department, University of
Chile, Beauchef 851, Santiago 8370456, Chile
2 Department of Statistical and Actuarial Sciences, The University of Western Ontario, 1151 Richmond
arXiv:2111.13666v2 [cs.SI] 16 Sep 2022
Street, London, Ontario, N6A 3K7, Canada.

4 Computational Intelligence Group, University of Basque Country, 20018 San Sebastián, Spain.
Abstract
Thin-file borrowers are customers for whom a creditworthiness assessment is uncertain due
to their lack of credit history. To address missing credit information, many researchers have
used borrowers’ social interactions as an alternative data source. Exploiting social networking
data has traditionally been achieved by hand-crafted feature engineering, but lately, graph
neural networks have emerged as a promising alternative. Here we introduce an information-
processing framework to improve credit scoring models by blending several methods of graph
representation learning: feature engineering, graph embeddings, and graph neural networks.
In this approach, we aggregate the methods’ outputs to be fed to a gradient boosting classifier
to produce a final creditworthiness score. We have validated this framework over a unique
multi-source dataset that characterizes the relationships, interactions, and credit history for
the entire population of a Latin American country, applying it to credit risk models, appli-
cation, and behavior. It also allows us to study both individuals and companies. Our results
show that the methods of graph representation learning should be used as complements;
they should not be seen as self-sufficient methods, as it is currently done. We improve the
creditworthiness assessment performance in terms of the measures of Area Under the ROC
Curve (AUC) and Kolmogorov-Smirnov (KS), outperforming traditional methods of exploit-
ing social interaction data. In the area of corporate lending, where the potential gain is much
higher, our results confirm that the evaluation of a thin-file company cannot solely consider
the company’s own characteristics. The business ecosystem in which these companies inter-
act with their owners, suppliers, customers, and other companies provides novel knowledge
that enables financial institutions to enhance their creditworthiness assessment. Our results
let us know when and on which population to use graph data and the expected effects on
performance. They also show the enormous value of graph data on the credit scoring problem
for thin-file borrowers, mainly to help companies with thin or no credit history to enter the
financial system.
Keywords: credit scoring; machine learning; social network analysis; network data; graph
neural networks
∗
NOTICE: This is a preprint of a published work. Changes resulting from the publishing process, such as editing, corrections,
structural formatting, and other quality control mechanisms may not be reflected in this version of the document. Please cite this work
as follows: Muñoz-Cancino, R., Bravo, C., Rı́os, S. A., & Graña, M. (2022). On the combination of graph data for assessing thin-file
borrowers’ creditworthiness. Expert Systems with Applications, 118809. DOI: https://doi.org/10.1016/j.eswa.2022.11880
1
1 Introduction
A large part of the population requires access to credit to achieve their life goals: social mobility,
owning a home, and financial success. Moreover, access to financial services and a proper credit
evaluation can facilitate and are often necessary to obtain a job, rent a home, buy a car, start a
new business, or pursue a college education (Aziz & Dowling, 2019; Hurley & Adebayo, 2017). At
the macroeconomic level, access to credit is a major driver for local economic growth, especially
in developing economies (Diallo & Al-Titi, 2017). Financial institutions play a significant social
role in facilitating access to credit and facing the entailed risks of lending money. To manage this
credit risk, financial institutions have applied credit scoring models to assess the creditworthiness
of their borrowers, that is, to distinguish between good and bad payers and delivering loans to
those who are most likely to repay. To build a credit scoring model, financial institutions often
use personal information, banking data, and payment history to estimate creditworthiness and the
probability of default. Despite being the standard mechanism in the industry for credit-granting
decisions and the management of the loan’s life cycle (Thomas, Crook, & Edelman, 2017), this
ubiquitous tool still does not ensure adequate access to credit and to the financial system.
The World Bank estimates that more than 1.4 billion adults remain unbanked, without access
to the financial system (The Global Financial Index, 2022). This number only considers those who
do not have a bank account through either a financial institution or mobile banking. If we included
underbanked people, that is, those who have an account but cannot apply for a loan, this number
would be much larger. Being unbanked or underbanked raises the issue of those who lack a credit
history, also known as thin-file borrowers: people who have no access to a loan not because they
are bad payers but because they lack the attributes evaluated by traditional credit scoring models
(Baidoo, 2020; Cusmano, 2018; Djeundje, Crook, Calabrese, & Hamid, 2021; Hurley & Adebayo,
2017).
In this scenario, lenders have tried different ways to reach this population; we highlight two
business models here. In the traditional business model, the higher risk assumed due to the
lack of information is compensated by applying higher interest rates. Alternatively, granting
microcredits has been used as a strategy to assess the client’s payment behavior under limited
exposure. However, neither of these solutions has proven to be cost-effective in addressing the
credit needs of this population (Baidoo, 2020; Hurley & Adebayo, 2017).
For this reason, financial institutions, fintech, and researchers have looked in recent years for
business-model innovations and better decision-making with the available information. This search
is done via developing better scoring algorithms and using alternative data sources to improve
credit scoring models. Regarding the use of alternative information, graph data has gained high
visibility because it allows an improvement of credit scoring models’ performance (Óskarsdóttir,
Bravo, Sarraute, Vanthienen, & Baesens, 2019; Roa, Rodrı́guez-Rey, Correa-Bahnsen, & Valencia,
2021).
We have identified two main gaps, which are addressed in this work. The first gap is the
data sources employed. Most of the studies are carried out with partial social networks that fail
to capture the overall picture of the client’s social interactions. These networks are limited by
the data provider. Our study uses social networking data to characterize the interactions of the
country’s entire population, encompassing the complete financial system. Secondly, the network
knowledge extraction is mainly done both through hand-made feature engineering (Freedman &
Jin, 2017; Niu, Ren, & Li, 2019; Óskarsdóttir et al., 2019; Ruiz, Gomes, Rodrigues, & Gama, 2017)
and, in recent years, through graph neural networks (Roa, Rodrı́guez-Rey, et al., 2021) that do are
no improvement over the traditional feature-engineering approach.
2
Our work will investigate the combination of different representation learning techniques with
complex graph structures instead of observing them in isolation. Hence, we formulate the following
research questions:
1. When combining different graph representation learning (GRL) techniques over complex
graph structures, is there a performance improvement compared to merely applying hand-
crafted feature engineering or graph neural networks?
2. What insights are obtained into the combined network features, and what value do these
insights add to credit risk assessment?
3. Where does social information help the most? Is the most significant performance enhance-
ment obtained in personal credit scoring or business credit scoring? What can we gather
from this information? Does it influence which network and which features are the most
relevant?
This study challenges traditional hand-crafted feature engineering and the novel approach of
graph neural networks (GNNs) by combining multiple GRL methods. In particular, our work
contributes to the following aspects.
• We introduce a framework to combine traditional hand-engineered features with graph em-

beddings and GNN features. This framework produces a single score, helping its users decide
whether to approve or reject a credit.
• Our results are the first to validate and test graph data regarding both corporate and con-
sumer lending, showing that the information from graphs has a different effect depending on
the analyzed borrower, people, or companies. These effects are reflected both in the predic-
tive power enhancement and in the features relevant in each problem, letting us know not
only when and on which population to use social-interaction data but also which effects on
creditworthiness prediction performance to expect.
• To the best of our knowledge, this is the first study that considers the credit behavior of
an entire country, together with social networks that allow the characterizing of its entire
population and consolidate multiple types of social and economic relationships, for example,
parental, spouses, business owners, employers and employees, or transactional services.
• This paper also contributes to the growing literature in credit scoring and network data,
proposing a mechanism to achieve better results than the popular hand-crafted feature en-
gineering and the novel GNN approach.
This paper is structured as follows. Section 2 presents a review of credit risk management,
credit scoring and social networks. The GRL methods are presented in Section 3. Section 4
describes the data sources and features extracted for classification. Section 5 shows the proposed
information-processing methodology and the adopted experimental design. Section 6 presents the
results obtained. The conclusions and future work that originated from this research are presented
in Section 7.
3
2 Background and Related Work
2.1 Credit Risk Management
Banks’ core business is granting loans to individuals and companies. Granting a loan is not
risk-free; in fact, banks are heavily exposed to credit risk (Anderson, 2022), originating from the
potential loss due to the debtors’ default or their inability to comply with the agreed conditions
(The Basel Committee on Banking Supervision, 2000). Banking risk management focuses on
detecting, measuring, reporting, and managing all sources of risk. Banks define strategies, policies,
and procedures to limit the assumed risk. These strategies encourage and integrate the use of
mathematical models for the early detection of potential risks. Credit scoring is widely used for
managing credit risk, handling large volumes of data, and capturing complex patterns that are
difficult to express as simple business rules. This instrument became popular and ubiquitous in the
1980s, mainly due to advances in computing power and to the growth of financial markets, which
made it almost impossible to manage large credit portfolios without this kind of tool (Thomas et
al., 2017).
The regulatory framework also endorses the use of credit scoring models; in fact, the Basel Ac-
cords allow banks to manage credit risk with internal ratings. Specifically, banks develop internal
models for assessing the expected loss. This assessment can be divided into three components: the
probability of default (PD), the loss given default (LGD), and the exposure at default (EAD). The
PD is a key component, because it is used to define the credit granting policies and for portfolio
management. The general approach to estimating the PD and assessing the borrower’s credit-
worthiness is through classification techniques using demographic features and payment history as
explanatory variables.
Over the years, lenders have explored multiple ways to improve creditworthiness assessment,
novel machine learning techniques (Moscato, Picariello, & Sperlı́, 2021), and non-traditional data
sources (Aziz & Dowling, 2019). Multiple lines of research have been established; some of them at-
tempt to understand the characteristics of defaulters (Bravo, Thomas, & Weber, 2015), the feature
selection process (Kozodoi, Lessmann, Papakonstantinou, Gatsoulis, & Baesens, 2019; Maldonado,
Pérez, & Bravo, 2017), or the transformation of the feature space (Carta, Ferreira, Reforgiato Re-
cupero, & Saia, 2021). However, the most significant improvements have been obtained by the
exploitation of alternative data sources such as telephone call data (Óskarsdóttir, Bravo, Vanathien,
& Baesens, 2018a; Óskarsdóttir et al., 2019, 2017), written risk assessments (Stevenson, Mues, &
Bravo, 2021), data generated by an app-based marketplace (Roa, Correa-Bahnsen, et al., 2021;
Roa, Rodrı́guez-Rey, et al., 2021), social media data (Cnudde et al., 2019; Putra, Joshi, Redi, &
Bozzon, 2020; Tan & Phan, 2018), network information (Ruiz et al., 2017), behavioral and psycho-
logical surveys (Goel & Rastogi, 2021), fund transfers datasets (Shumovskaia, Fedyanin, Sukharev,
Berestnev, & Panov, 2020; Sukharev, Shumovskaia, Fedyanin, Panov, & Berestnev, 2020), and psy-
chometric data (Djeundje et al., 2021; Rabecca, Atmaja, & Safitri, 2018; Rathi, Verma, Jain, Nay-
yar, & Thakur, 2022). All these studies have in common the use of social-interaction information,
the graph formed of the interactions among individuals recorded in alternative data sources.
There are multiple taxonomies of credit scoring problems. One that has been widely adopted
by academics and practitioners distinguishes between application scoring and behavior scoring.
On the one hand, application scoring corresponds to a credit scoring system for new customers,
where the available information is often scarce and limited. On the other hand, behavioral scoring
is a credit scoring system for borrowers with available credit and repayment history. In the current
study, both of these credit scoring types and their differences between personal and business clients
4
are explored.
2.2 Credit Risk and Social Networks

The inclusion of alternative data in the credit scoring problem has gained relevance in recent years.
We define graph data as any information that records the relationships or interactions among
entities that can be represented by a set of nodes in which edges connect pairs of nodes. We refer
to a network as a Social Network when nodes are people or companies, and edges denote any social
interaction, such as among friends, acquaintances, neighbors, colleagues, or affiliations with the
same group (Romero, Uzzi, & Kleinberg, 2019). Mathematically, we describe a network through
a graph G(V, E, A), where V is the set of nodes, and E is the set of edges. Let V = {v1 , . . . , vN }
where |V | = N is the number of nodes, and the adjacency matrix A ∈ R|V |×|V | with Aij = 1 if
there is an edge eij from vi to vj , Aij = 0 otherwise. Additionally, the graph can be associated
with a matrix of node attributes X ∈ RN ×F , where Xi ∈ RF represents the feature vector of node
vi .
The emerging literature on credit scoring and network data has focused on incorporating hand-
crafted features into a traditional credit scoring problem (Óskarsdóttir, Bravo, Verbeke, Baesens,
& Vanthienen, 2018; Óskarsdóttir et al., 2017). The authors incorporate network information into
the formulation of the customer churn problem, using eight telco datasets originating from around
the world. This series of studies outlines the foundations for the incorporation of network data
in credit scoring. This framework is applied to the credit scoring problem by Óskarsdóttir et al.
(2018a); Óskarsdóttir, Bravo, Vanathien, and Baesens (2018b); Óskarsdóttir et al. (2019), where the
authors introduce a methodology to enhance smartphone-based credit scoring models’ predictive
power through feature engineering from a pseudo-social network, combining social-network analysis
and representation learning. According to this research, it is feasible to increase the performance
of micro-lending smartphone applications, generating high helping potential for financial inclusion.
An extension of this work is to measure the temporal and topological dynamics of credit risk,
that is, how it evolves and spreads over the graph representation of the social network. For instance,
Bravo and Óskarsdóttir (2020) implemented modifications to the PageRank algorithm to quantify
this phenomenon. This methodology allows them to quantify the risk of the different entities in a
multilayer network. Their results show how the risk of default spreads and evolves over a network
of agricultural loans. Then, Óskarsdóttir and Bravo (2021) analyzed how to build the multilayer
network, interpret the variables derived from it, and incorporate this knowledge into credit risk
management. Their results reveal that the default risk increases as a debtor presents links with
many defaulters; however, this effect is mitigated by the size of each individual’s neighborhood.
These results are significant because they indicate that default and financial-stability risk spread
through the network.
Other works have used an approach based on a graph convolutional network (GCN) for this
purpose. Shumovskaia et al. (2020) present one of the first empirical works with massive graphs
created from transactions between clients of a large Russian bank. They propose a framework
to estimate links using SEAL (Zhang & Chen, 2018) and recurrent neural networks, the SEAL-
RNN framework. One of the advantages of using SEAL is that it focuses only on the link’s
neighborhood to be predicted, and it does not use the entire graph as in GCN. This framework
permits the analysis to be scaled to massive graphs of 86 million nodes and 4 billion edges. Although
the framework is not a methodology for default prediction, Shumovskaia et al. extend the scope of
their research and apply it to a credit scoring problem. Sukharev et al. (2020) propose a method to
predict the default from a money transfer network and the historical information of transactions.
5
To work with both datasets, they propose a methodology based on GCN and recurrent neural
networks to handle network data and transactional data, respectively. As baseline models, they
train a model with 7000 features; however, they achieve an increase of 0.4% AUC when comparing
the proposed model with the best baseline model. Finally, Roa, Correa-Bahnsen, et al. (2021)
present a methodology for using alternative information in a credit scoring model. Models are
estimated using data generated by an app-based marketplace. This information is precious for
low-income segments and young individuals, who are often not assessed well by traditional credit
scoring models. The authors compare a model with hand-crafted features versus models from
GCNs. However, GCNs do not achieve better results than do hand-crafted features in terms of
predictive power.
3 Representation Learning on Networks

The machine learning subfield that works on graph-structured data is known as graph represen-
tation learning or GRL (Hamilton, Ying, & Leskovec, 2017). Unlike the traditional tabular data,
network data imposes a challenge to conventional machine learning algorithms because it is not
possible to use them directly, forcing changes either on the algorithms or on the data representa-
tion. These challenges are required because the network information is, in essence, unstructured.
In fact, operations that are easy to calculate on other data types, such as convolutions on images,
cannot be applied directly to graphs because each node has a variable number of neighbors. Re-
searchers have proposed many methodologies to extract knowledge for networks; here, we present
a nomenclature and the characteristics of the most popular methods.
3.1 Feature Engineering

Data preparation is one of the most critical steps in any analytical project before training any
machine learning model. Formulating accurate and relevant features is critical to improving model
performance (Nargesian, Samulowitz, Khurana, Khalil, & Turaga, 2017). Regarding the use of
graph data, the traditional feature engineering approach consists of characterizing each node either
based on the aggregation of its neighborhood’s features or the node’s statistics within the network.
3.2 Network Embeddings

Network embedding methods are unsupervised learning techniques aiming to learn a Euclidean
representation of networks in a much lower dimension. Each node is mapped into a Euclidean
space through the optimization of similarity functions. The distance between network nodes in the
new space is a surrogate for the node’s closeness within the network structure. Node embedding
techniques often replace feature engineering processes.
Formally, a network is represented by a graph G(V, E, A) defined by a set of nodes V , a set
of edges E, and an adjacency matrix A ∈ R|V |×|V | . The embedding of a node is a function
f : G(V, E, A) → Rd that maps each node v ∈ V to an embedding vector {Zv }v∈V ∈ Rd (Arsov &
Mirceva, 2019), preserving the adjacency in the graph. The embedding vectors of pairs of nodes
that are connected by an edge are closer than those that are disconnected. Let Z ∈ R|V |×d denote
the node-embedding matrix, where d |V | for scalability purposes. The most popular network
embedding method is Node2vec (Grover & Leskovec, 2016), which is an algorithmic framework
for learning low-dimensional network representation. This algorithm maximizes the probability of
6
preserving the neighborhood of the nodes in the embedding subspace. The algorithm optimizes
using stochastic gradient descent, a network-based objective function, and produces samples for
neighborhoods of nodes through second-order random walks. The key feature of Node2vec is the
use of biased-random walks, providing a trade-off among two network search methods: breadth-
first search (BFS) and depth-first search (DFS). This trade-off creates more informative network
embeddings than other competing methods.
3.3 Graph Neural Networks (GNN)

Graph-structured data has arbitrary structures that can vary significantly between networks or
within different nodes of the same network. Their support domain is not a uniformly discretized
Euclidean space. For this reason, the convolution operator that is often used for signal processing
cannot be directly applied to graph-structured data. Geometric deep learning (GDL) and graph
neural networks (GNN) aim to modify, adapt and create deep learning techniques for non-Euclidean
data. The proposed GDL computational schemes are an adaptation of deep autoencoders, convo-
lutional networks, and recurrent networks to this particular data domain. In this study, we will
be applying graph convolutional networks (GCN) and graph autoencoders (GAE).
3.3.1 Graph Convolutional Networks

The Graph convolutional networks generalize the convolution operation to networks formalized
as graphs. The GCNs aim to produce a node’s representation Zv by adding its attributes or
feature vector Xv and neighbors {Xu }u∈N (v) , where N (v) is the ego network of node n, that is, the
subgraph composed of the nodes to whom node n is connected. This study uses the spectral-based
GCN, also known as the Chebyshev spectral convolutional neural network, proposed by Defferrard,
Bresson, and Vandergheynst (2016), which defines the graph convolution operator as a filter from
graph signal processing. In particular, we use the specific GCN proposed by Kipf and Welling
(2016a), which uses as a filter a first-order approximation of the Chebyshev polynomial of the
eigenvalues’ diagonal matrix. This graph convolution operator follows the expression:
1 1
Xi ∗ gθ = θ0 Xi − θ1 D− 2 AD− 2 Xi , (1)
where Xi is the feature vector, gθ is a function of the eigenvalues of
Pthe normalized graph, Laplacian
1 1
matrix L = In − D− 2 AD− 2 , A is the adjacency matrix, Dij = j Ai,j , ∀i, j ∈ V : i = j, and θ is
the vector of the Chebyshev coefficients. In the following, section we summarize the derivation of
this first-order approximation (Kipf & Welling, 2016a).
3.3.2 Derivation of GCN from Spectral Methods

Spectral methods are founded on a solid theoretical basis defined for methods of graph signal
processing developed essentially from the Laplacian matrix properties. To build up the graph
convolution operator, we start from the normalized graph Laplacian matrix defined as follows:
1 1
L = In − D− 2 AD− 2 , (2)
P
where A is the adjacency matrix, and Dij = j Ai,j , ∀i, j ∈ V : i = j. Because L is a real,
symmetric, and positive semi-defined matrix, we can rewrite L as a function of its eigenvector
matrix U and its eigenvalues λi , that is, L = U ΛU T , where U ∈ RN ×N , and Λij = λi , ∀i, j ∈ V :
i = j.
7
The next step is to define the graph Fourier transform and its inverse. The graph Fourier
transform F of the feature vector Xi ∈ X is defined as follows:
F(Xi ) = U T Xi , (3)
where X is the matrix of node attributes, and the inverse Fourier transform of a graph is defined
as follows:
F −1 (X̂i ) = U X̂i , (4)
where X̂i are the coordinates
P of the nodes in the new space. Therefore, the feature vector can
be written as Xi = jinV X̂i uj . Finally, the graph convolution of feature vector Xi with filter
g ∈ RN , using the element-wise product , is defined as follows:
Xi ∗ g = F −1 (F(Xi ) F(g)) (5)

One of the most popularPfilters is the Chebyshev polynomial of the eigenvalues’ diagonal matrix,
that is, gθ = diag(U T g) = K θk Tk (Λ̂), where Λ̂ = 2λ/λmax −I and the polynomials Tk are defined
as Tk (x) = 2xTk−1 − Tk−2 (x), with T0 (x) = 1 and T1 (x) = x. Therefore, the GCN, defined as the
Chebyshev Spectral CNN (Defferrard et al., 2016), takes the following form:
X
Xi ∗ gθ = θk Tk (L̂)Xi , (6)
K
where L̂ = 2L/λmax − I. Despite being a graph convolution simplification, this convolution

is computationally expensive for large graphs. To solve this problem, Kipf and Welling (2016a)
present a first-order approximation of the Chebyshev Spectral CNN. Assuming K = 1 and λmax = 2
, the Equation 5 takes the following form:
1 1
Xi ∗ gθ = θ0 Xi − θ1 D− 2 AD− 2 Xi (7)
3.3.3 Graph Autoencoders (GAEs)

Graph autoencoders (GAEs) are an unsupervised method to obtain a low-dimensional represen-
tation of the network. The objective of the GAE is to reconstruct the original network using the
same network for this task but encoding it, reducing its dimensionality, and then decoding it to
reconstruct the network. The encoded representation is used as the network embedding. Wu et
al. (2020) distinguish two main uses of GAEs, namely graph generation and network embedding.
This research will use GAEs to obtain a lower-dimensional vector, preserving the network topology
(network embedding).
The previously defined GCN is the building block of the GAE architecture and allows the
simultaneous encoding of the network topology and the attributes of the nodes. The GAE (Kipf &
Welling, 2016b) calculates the network embedding matrix Z and the reconstruction of the original
network adjacency matrix Â as follows:
Â = σ(ZZ T ), with Z = GCN (X, A), (8)

where X is the matrix of node attributes, and A is the network’s adjacency matrix.
8
4 Data Description
The data used in this paper encompasses several datasets provided by a large Latin American
bank. Some datasets contain information from their customers, while others concern the entire
population of the country.
4.1 Ethical and Privacy Protection Considerations

The datasets contain anonymized information and do not compromise the identity of any customer
or their personal information in any way. The datasets share an anonymized customer’s ID allowing
us to merge multiple sources. Regarding the value, importance, and sensitivity of the data, we have
applied multiple actions to ensure its security, integrity, and confidentiality. Customer identifiers
and any personal data were removed before starting the analysis, and there is no possibility that
this investigation can leak any personal private information. In addition, any final data produced
as a result of this research does not compromise customers’ privacy.
4.2 Social-Interaction Data

The information collected by the financial institution to construct the social network background
information of the thin file borrower originates from varied sources and can be cataloged as follows:
• [WeddNet] Network of marriages: This network is built from the information of mar-
riages recorded by the bureau of vital statistics from 1938 to December 2015. It includes the
anonymized identifiers of the husband and the wife and the wedding date.
• [TrxSNet] Transactional services network: The primary source of this network comes
from transactional services data, primarily payroll services and the transfers of funds between
two entities. We have access to monthly data from January 2017 to December 2019.
• [EnOwNet] Enterprise’s ownership network: This network is built from the informa-
tion on companies’ ownership structure. For each firm, we have information concerning their
owners, be they individuals or other firms. We have quarterly information from January
2017 to November 2019.
• [PChNet] Parents and children network: This network corresponds to parental relation-
ships. For every person born between January 1930 and June 2018, we have the anonymized
identifiers of their parents.
• [EmpNet] Employment network: This network is built from multiple sources and con-
nects people with their employers. We have monthly data from January 2017 to December
2019.
4.3 Financial Data

The node dataset contains information on the consolidated indebtedness of each debtor in the
financial system from January 2018 until March 2020, reporting monthly the debt decomposition
from 7.65 million people and 245,000 firms. We refer to the features extracted from this dataset
as node features. Additionally, for every person and firm in the previous dataset, we have access
to the BenchScore, which corresponds to the probability of default for the coming 12 months.
9
This probability was assessed and provided by the financial institution, and it is our benchmark
to contrast the performance of our models.
4.4 Network Construction

It is possible to build a network from each of the data sources indicated in Section 4.2. However,
they share characteristics that allow them to be grouped. For this reason, we combine the networks
into two primary data sources, from which we construct networks that characterize people and
businesses.
[FamilyNet] Family network: This network is formed through the combination of the net-
work of marriages (WeddNet) and the parent and children network (PChNet). For the construction
of this network, we use the historical information available until the beginning of the analysis pe-
riod; no further information is included. In this way, the network remains unvarying throughout
the study. We call this type of network a static network.
[EOWNet] Enterprise’s ownership Network and Workers: This network is composed
of the fusion of the transactional services network (TrxSNet), the enterprise’s ownership network
(EnOwNet), and the employment network (EmpNet). This network attempts to represent the
business ecosystem in which companies, business owners, and employees interact. Based on these
data sources, a series of 24 networks are generated, one for each of the 24 months available,
including the information collected up to the last day of the corresponding month. We call this
type of network a temporal network, because nodes and edges change over time.
5 Experimental Design and Methodology

5.1 Datasets
The credit scoring models are built with information about the financial system for 24 months.
However, the models are trained over 23 months because the first month is left out for the feature
extraction process (presented in Section 5.3) in order to avoid target leakage (Kaufman, Rosset,
Perlich, & Stitelman, 2012). For the unbanked application scoring model, individuals and com-
panies are considered only in the month that they enter the financial system. In contrast, for
the behavioral scoring model individuals and companies are considered six or more months after
entering the financial system. Table 1 summarizes the scoring application, the model trained and
the size of the dataset used.
Table 1: Description of dataset
Scoring application Model Observations # Features
Unbanked Application Business Credit Score 29,044 687
Scoring Personal Credit Score 192,942 1,283
Business Credit Score 931,910 687
Behavioral Scoring
Personal Credit Score 1,978,664 1,283
5.2 Target
The target event was ”becoming a defaulter during the period of observation”. Therefore, we only
took into account individuals or businesses that were non-defaulters at the start of the period of
observation; we dismissed entities that were defaulters at the very beginning of the observation.
10
In the current study, a person or company was considered a defaulter when they had payments
past the due date for 90 or more days within 12 months starting from the observation point.
Otherwise, they were considered non-defaulters. The target vector, denoted by ydef , contained the
actual information about the target event.
5.3 Traditional and Graph Representation Learning Features

Combining the node information from Section 4.3 and the network data from Section 4.2 makes it
possible to generate a set of new characteristics through a feature extraction process. The sets of
characteristics generated are detailed below:
• [NodeStats] Node Statistics: This dataset collects node centrality statistics, namely, its
degree, degree centrality, number of triads, PageRank score, authority and hub score given by
the Hits algorithm (Kleinberg, 1999), and an indicator of whether the node is an articulation
point.
• [EgoNet] EgoNetwork Agreggation: In this dataset, each node is characterized by the

information of other nodes connected to it (ego network). We refer to the dataset as egoNet
aggregation features when we apply some aggregation function to the characteristics of
the nodes included in the ego network. Specifically, for each attribute in the NodeStats
feature set we apply the mean and SD in this study as in Nargesian et al. (2017); Roa, Correa-
Bahnsen, et al. (2021). Figure 1 shows how this process would be within the network; the
black node corresponds to the target node, and the gray nodes belong to its Ego Network.
Computing the egoNet aggregation features assumes that each connection in the ego network
has the same importance; however, connections in ego networks can be weighted. For this
reason, we compute the egoNet weighted aggregation features where the features of the
neighboring nodes are weighted according to a measure of the relationship intensity measured
by the weighted average and SD of the NodeStats attributes.
Figure 1: Example of Network Features
• [N2V] Node2Vec Features: Node2Vec is an unsupervised method that only uses the
network structure to generate the graph embedding. For the static network FamilyNet,
Node2Vec is applied only once. A node is characterized by this embedding regardless of the
moment it was sampled in the dataset. For temporal networks (EOWNet), Node2Vec has
to be recomputed every period because of the network changes. Each node is characterized
by the embedding corresponding to the month in which it was sampled in the dataset.
11
t0 tn time
train train train

f1 (G1 ) f2 (G2 ) fn (Gn )
... ...
{Zv }v ∈ V1 {Zv }v ∈ V2 {Zv }v ∈ Vn
sample sample sample sample
dataset
Figure 2: Node2Vec to Features
t0 tn time
train model apply apply apply

f (Go ) f (G1 ) f (G2 ) f (Gn )
... ...
{Zv }v ∈ V1 {Zv }v ∈ V2 {Zv }v ∈ Vn
sample sample sample sample
dataset
Figure 3: Graph Convolutional Networks and Graph Autoencoders to Features
Figure 2 shows the process through which to obtain the embedding features by applying
Node2Vec. Each period, a new model is trained, and the resulting embedding is consolidated
to characterize the sample dataset that will allow us to train the final credit scoring models.
• [GNN] Graph Neural Network Features: We used either GCN or GAE for the ex-
traction of GNN features. These methods carry out the graph convolution of the network
structure through a node feature vector in order to generate as output the feature vector
for ensuing classification by machine learning approaches. In this study, GNN input feature
vectors are in the Node feature dataset. Regardless of whether the network is static or
temporal, the feature vector is dynamic and changes during each observation period, that is,
each month. To avoid target leakage (Kaufman et al., 2012), we train the GCNs with the first
available data period; these models are applied in the subsequent months while the network
or the feature vectors change. This approach does not handle new entrants to the social
interaction networks; however, due to the sources of social interaction information employed
in this study, most thin-file borrowers are taken into account in the training dataset, despite
having temporal networks. Therefore, new entrants do not affect our results. Under other
circumstances, for instance when working with a partial network, our recommendation is to
train a new model for every period, taking precautions not to incur in target leakage on the
population of the credit scoring model. Depending on the dynamism of the network, one
12
solution is to calculate local connection updates as suggested in Vlasselaer et al. (2015) in the
interim between model training phases. The GCNs are trained on the entire network (either
FamilyNet or EOWNet) regardless of whether the nodes belong to our training dataset or
not. The output of the GCN is the label of the nodes in the network, which can be ei-
ther defaulter, non-defaulter, or unbanked; therefore, each GCN solves a multi-classification
problem by providing the a posteriori probabilities of each label for each node. Finally, each
node is characterized by the GCN Features resulting from the application of the GCN on
the network and on the feature vector of the month in which the node was sampled, that is,
in which it entered the banking system. Figure 3 illustrates the feature-engineering process
to extract the GCN features. Regarding GAE, the feature engineering process is similar. In
this case, the embedding corresponds to the bottleneck hidden layer representing the encoder
section’s output. For consistency, we apply the same data selection applied for the GCN to
the training of the GAE models, although their training is unsupervised.
5.4 Feature Subsets

The datasets for machine learning training and validation experiments are composed of the follow-
ing subsets of features.
• Subset A : XN ode is the dataset of node characteristics.
• Subset B : XBenchScore is the benchmark score; this attribute is also used as a benchmark to
quantify the performance of our proposed approach.
• Subset C : XN odeStats is composed of the statistics obtained from the position of the node
within the network.
• Subset D : XEgoN et includes the egoNet aggregation and the egoNet weighted aggregation
features that are calculated in three scenarios, considering the entire network, considering
only those edges that are bridges, and considering those edges that are not bridges∗ .
• Subset E : XGN N,N 2V corresponds to the features created by applying GNNs and Node2Vec.
People are characterized by features from both networks (EOWNet and FamilyNet), while
companies are characterized only by EOWNet.
For this study, we aggregated these feature subsets into eight increasingly larger datasets for
training and validation, each defining a different experimental setting. The details of the feature
sets are presented below in Table 2.
5.5 Evaluation Metrics

The AUC is a popular metric used to evaluate the model performance in classification problems
(Zeng & Zeng, 2019). It ranges between 0.5 and 1. Values closer to 1 indicate a better discrimina-
tory capacity, while a value of 0.5 indicates a performance equivalent to a chance decision. In the
context of credit scoring, the AUC can be easily interpreted as follows: For a randomly selected
defaulter and non-defaulter pair, the AUC corresponds to the probability that the classification
model assigns a higher score to the defaulter.
∗
An edge connecting the nodes u and v is called a bridge; if removing this edge, there is no longer a path
connecting u and v
13
Table 2: Experiments Setup
Experiment Id Feature Group
A X = {XN ode }
A+B X = {XN ode + XBenchScore }
A+B+C X = {XN ode + XBenchScore + XN odeStats }
A+B+D X = {XN ode + XBenchScore + XEgoN et }
A+B+E X = {XN ode + XBenchScore + XGN N,N 2V }
A+B+C+D X = {XN ode + XBenchScore + XN odeStats + XEgoN et }
A+B+C+E X = {XN ode + XBenchScore + XN odeStats + XGN N,N 2V }
A+B+C+D+E X = {XN ode + XBenchScore + XN odeStats + XEgoN et + XGN N,N 2V }
Another extensively utilized performance measure is the Kolmogorov-Smirnov (KS) statistic,

which measures the distance separating the cumulative distributions of defaulters (PD (t)) and
non-defaulters (PN D (t)) (Fang & Chen, 2019). The KS statistic is defined as:
KS = max |PD (t) − PN D (t)| (9)

t
and KS ranges between 0 and 1, and a higher KS indicates a higher prediction performance.
5.6 Methodology
First, we carry out a feature engineering process that seeks to create attributes to characterize
the nodes from the network. Then, in the train-test split step, the available dataset is divided
into a training dataset of which 30% consists of the samples used to estimate the model’s hyper-
parameters, and the remaining 70% consists of the samples used to train and validate models
according to an N-Fold Cross-Validation scheme.
Before estimating the best hyper-parameters, we apply a feature selection process. In this step,
the intention is to choose a low-correlated subset of features with high predictive power. Three
selection levels are formulated. The first is a bivariate selection that only considers one feature
at a time and the target vector to build a prediction model, evaluating its performance. The
selection process applies a threshold to this feature’s predictive power. Only those variables are
selected such that KS > KSmin and AU C > AU Cmin , where KSmin and AU Cmin are threshold
parameters.
Next, a multivariate selection is applied following a simple but effective method to drop cor-
related features that we have developed. This algorithm starts with an empty list S. We iterate
over a set of features P in decreasing order of predictive power and append to S those features
whose absolute value of the correlation with each and all the features in S is less than a threshold
ρ set to avoid high correlated features (Akoglu, 2018). The algorithm stops when all the features
have been visited. The first feature in P is added to S without correlation comparison.
For this study, the multivariate selection process is applied twice. In the first application,
the process selects low-correlated features for each group of attributes P ∈ {XN ode ∪ XBenchScore ,
XN odeStats , XEgoN et , XGN N,N 2V }. Secondly, it is applied globally to all the remaining features P =
{XN ode ∪ XBenchScore ∪ XN odeStats ∪ XEgoN et ∪ XGN N,N 2V }. In both cases, a threshold ρ is used, and
the features are ordered by the features’ AUC, from higher to lower.
Finally, at the N-Fold Cross-Validation stage, the dataset is partitioned into N subsets of equal
size. Each subset is used alternatively as the test dataset, while the remaining folds are used to
train the classification model. The hyper-parameters used in each iteration are those estimated
14
in the previous stage. Additionally, in each of these iterations, multiple models are trained with
different feature sets and stored to be used later to compare the models.
5.7 Experimental Setup

The parameters of the univariate selection are set at KSmin = 0.01 and AU Cmin = 0.53; for the
multivariate selection process, ρ = 0.7 in both processes to avoid high correlated features (Akoglu,
2018). The N-Fold Cross-Validation stage is carried out considering N = 10, and in each iteration,
the results of gradient boosting (Friedman, 2001) models are displayed. Other classification models
such as regularized logistic regression and Random Forest (Breiman, 2001) were trained. However,
gradient boosting consistently delivered better results.
6 Results and Discussion

In this section, we present the results obtained. We begin with the technical implementation
details; then, we analyze the execution times. Subsequently, we display the model’s performance
in three scenarios: the impact on performance using traditional features, the impact on performance
using the different graph representation methods, and the advantages of combining these methods.
Finally, an analysis is presented of the main features, traditional and network-based, for the
creditworthiness assessment.
6.1 Implementation Details

In this work, we used the Python implementations Networkx v2.6.3 (Hagberg, Swart, & SChult,
2008) and Stanford Network Analysis Platform (SNAP) v5.0.0 (Leskovec & Sosič, 2016) in the
hand-crafted feature engineering process (XN odeStats , XEgoN et ); for Node2Vec, GCN and GAE
(XGN N,N 2V ) we used PyTorch v1.6.0 (Paszke et al., 2019) and PyTorch Geometric v2.0.1 (Fey &
Lenssen, 2019).
To conduct the experiments, we used a laptop with an Intel 8-Core i7 CPU and 32 GB of RAM
for network construction and hand-crafted feature engineering. For the Node2Vec, GCN, GAE,
and model training phases, we used a server with a driver node with 140 GB of RAM and 20 CPU
cores and between two and eight auto-scaling worker nodes with 112GB of RAM and 16 CPU
cores.
6.2 Execution Time

Below we detail the execution time of the most critical stages of our work, the implementation of
the GRL methods, and the models’ training.
• [NodeStats] Node Statistics: This process corresponds to the computation of the metrics
defined in Section 5.3. This process is carried out only once for the static network FamilyNet,
and it is calculated for all the available periods (24) of the EOWNet. The computation of
all the metrics for a network takes, on average, 25 minutes. The total execution time of this
stage was 625 minutes.
• [EgoNet] EgoNetwork Aggregation: This process is calculated once per network type
(FamilyNet and EOWNet). The total execution time of this stage was 300 minutes.
15
• [N2V] Node2Vec Features: This process is carried out only once for the static network
FamilyNet, and it is calculated for all the available periods (24) of the EOWNet. The
Node2Vec training for a network takes, on average, 300 minutes. The total execution time
of this stage was 7,500 minutes.
• [GNN] Graph Neural Network Features: In this stage, eight models are trained using
each network and eight different feature vectors. The training of each model is carried out
over the first available period data; afterward, the trained model is applied to the data of
the remaining periods. Table 3 shows the execution time by GNN type and by network type.
The total execution time of this stage was 6,920 minutes.
Table 3: GCN and GAE Execution Time

Total Total
Unit Train
GNN Type Network Type Training Apply
Time (min)
Time (min) Time (min)
EOWNet 25 200 600
GCN
FamilyNet 60 480 600
EOWNet 140 1,120 1,200
GAE
FamilyNet 190 1.520 1.200
Total Execution Time (min) 3,320 3,600
• Gradient boosting training: Four models were trained using the methodology described
in Section 5.6 for the scenarios defined in Section 4, predicting application and behavioral
scoring for individuals and companies. The complete training for each scenario takes, on
average, 40 minutes. The total execution time of this step was 160 minutes.
The total execution time of the application of the proposed methodology to the datasets is
15,500 minutes. Although a large part of these executions was parallelized using the server de-
scribed in Section 6.1, the high computational cost is mainly due to two factors, the volume of
data and the complexity of the algorithms. Regarding the large volume of data, the FamilyNet
has 20 million nodes and 30 million edges, and the EOWNet has 8.6 million nodes and 26 million
edges. These massive dataset sizes directly impact the algorithms used, because the complexity
depends on the nodes |V |, edges |E| and the embedding dimension d; the algorithmic complexity
for this particular case is the following: Node2Vec: O(|V |d) (Grover & Leskovec, 2016), Graph
Convolutional Networks: O(|E|d) (Kipf & Welling, 2016a) and Graph Autoencoder: O(|V |2 d)
(Kipf & Welling, 2016b).
6.3 Model Performance Results

The model training process was conducted for each Experiment ID using different Feature Set
combinations as specified in Section 5.1. The results are presented in Tables 4 and 5. These
tables show the relative improvement in AUC and KS achieved by each model over the baseline
C −BenchScoreAU C −BenchScoreKS
BenchScore, measured as rowAUBenchScore AU C
and rowKS
BenchScoreKS
, respectively. Each row cor-
responds to an experiment defined in Section 5.4, and each column displays the four credit scoring
scenarios illustrated in Section 5.1.
The reported results correspond to the average of 10-fold cross-validation, as indicated in
Section 5. A t-test is applied to establish the statistical significance of the performance differences
obtained using the different feature sets, according to the suggestions proposed by Flach (2012).
16
Table 4: Improvement in AUC relative to the benchmark model (mean and std). We only report results
when the equal performance hypothesis is rejected, with a confidence level of 95%; otherwise, we display
*. The best performance in each column is shown in bold; more than one bold value indicates that the
hypothesis of equal performance between those models cannot be rejected.
Feature Business Credit Score Personal Credit Score
Set Application Behavioral Application Behavioral
A -3.52% ± 2.87% -0.90% ± 0.21% -0.74% ± 0.63% -0.63% ± 0.09%
A+B * 0.58% ± 0.06% 1.45% ± 0.39% 0.95% ± 0.06%
A+B+C * 1.13% ± 0.12% 2.02% ± 0.49% 1.07% ± 0.06%
A+B+D 8.96% ± 3.37% 2.33% ± 0.15% 2.31% ± 0.64% 1.25% ± 0.08%
A+B+E 3.92% ± 2.03% 1.77% ± 0.13% 3.17% ± 0.55% 1.96% ± 0.04%
A+B+C+D 9.00% ± 3.47% 2.37% ± 0.16% 2.39% ± 0.60% 1.32% ± 0.08%
A+B+C+E 4.25% ± 1.84% 1.94% ± 0.16% 3.26% ± 0.48% 2.03% ± 0.05%
A+B+C+D+E 8.43% ± 2.83% 2.80% ± 0.16% 3.58% ± 0.61% 2.18% ± 0.04%
Table 5: Improvement in KS relative to the benchmark model (mean and std). We only report results
when the equal performance hypothesis is rejected, with a confidence level of 95%; otherwise, we display
*. The best performance in each column is shown in bold; more than one bold value indicates that the
hypothesis of equal performance between those models cannot be rejected.
Feature Business Credit Score Personal Credit Score
Set Application Behavioral Application Behavioral
A * -4.15% ± 0.94% -5.25% ± 2.40% -2.39% ± 0.46%
A+B * 1.56% ± 0.40% 4.38% ± 1.19% 1.95% ± 0.35%
A+B+C * 3.21% ± 0.71% 6.27% ± 1.02% 2.23% ± 0.39%
A+B+D 20.69% ± 16.73% 7.69% ± 0.92% 6.79% ± 1.36% 2.69% ± 0.47%
A+B+E 12.22% ± 10.89% 5.83% ± 0.74% 8.64% ± 2.13% 4.68% ± 0.28%
A+B+C+D 21.28% ± 17.10% 8.09% ± 0.95% 7.12% ± 1.52% 2.83% ± 0.52%
A+B+C+E 12.88% ± 10.11% 6.33% ± 0.70% 8.93% ± 1.98% 4.93% ± 0.26%
A+B+C+D+E 19.32% ± 14.77% 9.45% ± 0.85% 10.83% ± 1.98% 5.15% ± 0.42%
6.3.1 Model Performance Using Traditional Features

The first step toward reaching a conclusion on the contribution of network data is to understand
whether our methodology enables us to obtain equal or better results than the current decision-
making scheme in the financial institution. For this, we compare the BenchScore with the results
obtained by the feature set A+B. A comparison with only the feature set A is not entirely accurate,
considering that we do not have access to all the features used in the BenchScore training.
The results show that our methodology obtains equal or greater performance, measured in
terms of AUC and KS statistics, in all four scenarios; three of them are greater, with statistically
significant differences.
The performance enhancements in Behavioral Business Credit Scoring are 0.58% and 1.56% for
AUC and KS, respectively. In terms of Personal Credit Scoring, the performance enhancements
are AUC: 1.45%, KS: 4.38% for Application Scoring and AUC: 0.95%, KS: 1.95% for Behavioral
Scoring. These results indicate that using our methodology and training a model with similar
features to the benchmark model can obtain better results than the current decision scheme applied
by the financial institution.
6.3.2 Model Performance Using Graph Representation Learning Features

In Application Business Credit Scoring, a nearly 9% AUC increase over the BenchScore is achieved.
These results are obtained by a model that incorporates three feature sets: A+B+D, A+B+C+D,
and A+B+C+D+E. Note that in these three sets, the common attributes correspond to the
traditional features A+B : XN ode ∪ XBenchScore and EgoNet Aggregation Features D : XEgoN et .
17
The performance comparison between all feature sets is shown in Table 6; we marked as * those
comparisons with no statistically significant differences.
Table 6: Performance comparison of Business Application Scoring. Performance is measured by the

C −columnAU C
relative increase in AUC ( rowAU
columnAU C ).
BENCH A A+B A+B+C A+B+D A+B+E A+B+C+D A+B+C+E A+B+C+D+E
BENCH * 3.65% * * -8.23% -3.77% -8.26% -4.08% -7.77%
A -3.52% * -2.92% -4.55% -11.45% -7.16% -11.49% -7.45% -11.01%
A+B * 3.01% * -1.68% -8.79% -4.37% -8.83% -4.67% -8.34%
A+B+C * 4.77% 1.71% * -7.23% -2.73% -7.26% -3.04% -6.77%
A+B+D 8.96% 12.94% 9.64% 7.79% * 4.85% * 4.52% *
A+B+E 3.92% 7.71% 4.57% 2.81% -4.63% * -4.66% * -4.15%
A+B+C+D 9.00% 12.98% 9.68% 7.83% 4.89% * 4.56% *
A+B+C+E 4.25% 8.05% 4.90% 3.13% -4.32% * -4.36% * -3.85%
A+B+C+D+E 8.43% 12.38% 9.10% 7.26% * 4.33% * 4.00% *
When performance is measured in terms of KS, the maximum is obtained with five feature sets,
with GRL features, but no method for feature extraction predominates. Although differences are
observed in the values presented, these are not statistically significant. From these results, it is
necessary to highlight at least one GRL method in the best feature sets. The complete comparison
is presented in Table 7
Table 7: Performance comparison of Business Application Scoring. Performance is measured by the

−columnKS
relative increase in KS ( rowKS
columnKS ).
BENCH A A+B A+B+C A+B+D A+B+E A+B+C+D A+B+C+E A+B+C+D+E
BENCH * * * * -17.14% -10.89% -17.55% -11.41% -16.19%
A * * * * -19.94% -13.90% -20.34% -14.40% -19.02%
A+B * * * * -16.16% -9.83% -16.57% -10.36% -15.20%
A+B+C * * * * -16.02% -9.68% -16.43% -10.21% -15.06%
A+B+D 20.69% 24.91% 19.28% 19.08% * * * * *
A+B+E 12.22% 16.14% 10.90% 10.72% * * * * *
A+B+C+D 21.28% 25.53% 19.86% 19.67% * * * * *
A+B+C+E 12.88% 16.83% 11.56% 11.37% * * * * *
A+B+C+D+E 19.32% 23.49% 17.92% 17.73% * * * * *
The best performance is observed in other scenarios when combining traditional features and
all the GRL features; this corresponds to the Feature Set A+B+C+D+E. The best performance
is achieved in AUC (see Table 4) and KS (see Table 5).
These results are of great importance because they indicate that the methods combined by our
methodology are complementary, and none is significantly better than the others. Both methods,
namely hand-crafted feature engineering and GNNs, have until now been treated in the literature
as independent in addressing the credit scoring problem.
When comparing the results of Application and Behavioral Credit Scoring, it is observed that
the most significant increase in performance, regardless of the metric, is achieved in Application
Credit Scoring. Network-related features complement the least availability of information, such
that the relationships that a person or company has are relevant when predicting their creditworthi-
ness. These results are of high interest for lenders and in terms of their strategies for the unbanked.
The improvement in predictive performance implies that more borrowers can be serviced without
increasing the portfolio default rate.
Regarding Behavioral Credit Scoring, traditional attributes are already good predictors of
creditworthiness; the borrower’s credit behavior is a good indicator of default. For this reason, the
increase in predictive performance is more limited, although still significant.
18
6.3.3 The Advantages of Blending Graph Representation Learning
The previous sections have shown that our approach allows us to enhance the discrimination
power of our benchmark in terms of AUC and KS. Through the incorporation of the graph data
by means of the GRL methods, this increase is even more significant. Now, we are interested in
discovering the contribution of each of these methods. The performance comparison between the
A+B+C+D+E feature set and each method by itself is shown in Tables 8 and 9, for AUC and
KS respectively; we marked as * those comparisons with no statistically significant differences. In
each table, the results are presented for each credit scoring scenario and the comparison using the
XEgoN et (A + B + D) and XGN N,N 2V (A + B + E) features; In both cases, the models trained
with the XN odeStats features are also included.
Table 8: Blended Graph Representation Learning performance. The performance enhancement of training
a model using all GRL methods (A+B+C+D+E) is measured as the relative increase in AUC given by
AU C −columnAU C
( [A+B+C+D+E]
columnAU C ).
Feature Set
Scoring Model
A+B+D A+B+C+D A+B+E A+B+C+E
Application Business Credit Score * * 4.33% 4.00%
Scoring Personal Credit Score 1.23% 1.16% 0.39% 0.31%
Behavioral Business Credit Score 0.47% 0.43% 1.02% 0.85%
Table 9: Blended Graph Representation Learning performance. The performance enhancement of

training a model using all GRL methods (A+B+C+D+E) is measured as the relative increase in KS
KS −columnKS
( [A+B+C+D+E]
columnKS ).
Feature Set
Scoring Model
A+B+D A+B+C+D A+B+E A+B+C+E
Application Business Credit Score * * * *
Behavioral Business Credit Score 1.68% 1.31% 3.47% 2.99%
The results show that combining the GRL methods always generates better or similar results
than using each method independently. An equal performance is only obtained for the Business
Application Credit Score, where the only statistically significant increase, in AUC terms, occurs
when using the XGN N,N 2V features. However, this feature subset does not produce an increment
compared to using only the XEgoN et features. On the other hand, in all other scenarios, the GRL
combination generates statistically significant increments, independent of the method used and
whether or not the XN odeStats features are incorporated. In this way, our approach allows us to
increase discriminatory power in assessing creditworthiness, generate more accurate models, and
use graph data better through a framework that combines multiple methods of GRL.
6.4 Feature Importance Analysis

To determine the importance of each feature, we utilize SHAP: SHapley Additive exPlanations
(Lundberg & Lee, 2017), an approach based on game theory that calculates each attribute’s im-
portance by comparing the model predictions with and without the attribute. The global feature
importance is examined with regard to the four scenarios described earlier, that is, the prediction
19
of Application or Behavioral Scoring for individuals or for businesses. All analyses are conducted
with the feature set A+B+C+D+E, which incorporates all the features and is the one that
reports the best results.
6.4.1 Business Credit Scoring

In Figure 4, the importance of the attributes incorporated into the model is presented. Figures 4(a)
and 4(c) show each attribute’s importance for Application and Behavioral Scoring, respectively;
we define it as the average of the absolute values of the SHAP values. Figures 4(b) and 4(d) show
each feature’s impact on the model output; only the 15 most relevant attributes are displayed in
both figures. Figures 4(b) and 4(d) allow us to understand how the value of a particular feature
affects the probability of default.
Figure 4: Business Credit Scoring: Feature Importance
(a) Application: Average impact on model output (b) Application: Impact on model output
(c) Behavioral: Average impact on model output (d) Behavioral: Impact on model output
As expected, the most significant contribution to the model is the BenchScore, which already
summarizes valuable information about each company that allows the estimation of its credit-
worthiness. This influence occurs in both scenarios, Application, and Behavioral. However, its
importance is more significant in Behavioral Scoring.
Among these 15 relevant attributes in both scenarios, only the BenchScore, commercial debt
amount (NODEATT 08), and unused revolving credit amount (NODEATT 05) correspond to
the business-related characteristics. The remaining top features correspond to Network-related
features.
An additional relevant feature is the average BenchScore of the company’s ego network, in-
cluding only the non-bridge edges. This result indicates the creditworthiness of the company’s
neighborhood is also highly predictive of the company’s creditworthiness. In Application Scoring,
20
this feature is practically as relevant as the BenchScore. See Table A1 for more detail on the
description of the most relevant variables. This table presents the taxonomy of the features used
in the current study, giving the necessary specifications for the correct interpretation of the feature
attributes and the nomenclature used for the management of the datasets.
Further, we find attributes whose influence corresponds to people related to the company,
including its owners, for instance, the attribute generated from a Graph Autoencoder trained with
the consumer debt of the EOWNet Network. The consumer-debt effect of the ego network is
also observed in the attribute corresponding to the consumer debt weighted by the PageRank of
the node’s neighborhood. The presence of consumer debt as a relevant network-related feature
in Business Scoring is highly significant, especially in SMEs. The EgoNet’s short-term personal
debt, mainly on the part of the owners, accounts for the often blurred separation between personal
finances and company finances. The owner’s default can affect the company and vice versa. This
hypothesis requires a more detailed investigation and will be addressed for future work.
To quantify the usefulness, impact and importance of the different feature sets on the output
model, Figure 5 presents a Treemap based on the average of the absolute values of the SHAP
values; the complete list of the model features is displayed, and the different colors indicate that
they belong to different feature sets.
Figure 5: Business Credit Scoring: Treemap of Feature Importance, the Average Impact on Model Output
(a) Application Scoring
(b) Behavioral Scoring
In Figure 5(a), it is shown that the feature set XEgoN et (D) contributes, in Application Scoring,
60% of the model’s overall impact. The feature set XGN N,N 2V (E) contributes 21%, of which 19%
correspond to GNN features, while 2% correspond to Node2Vec. The low importance of Node2Vec
features is likely the reason for the limited research on Node2Vec to enhance the prediction of
creditworthiness.
However, in Business Behavioral Scoring, the traditional characteristics now represent 48 % of
the total impact of the model. In contrast, in Business Application Scoring, they represent only
21
16% (See Figure 5(b)). Indeed, the BenchScore attribute alone represents 29% of the total impact.
The feature set XEgoN et (D) represents 30% of the total impact, the average BenchScore of the
ego network being the most relevant attribute.
6.4.2 Personal Credit Scoring

In Personal Scoring, the person’s characteristics (XN ode + XBenchScore ) produce a more meaningful
impact than the business score. The person’s attributes represent 37% and 47% of the total impact
for Application and Behavioral Scoring, respectively.
Besides the BenchScore, other relevant features are the amount of consumer debt (NODEATT 07)
and amount of unused revolving credit, and total debt amount (NODEATT 01) (see Figures 6(a)
and 6(c)).
Figure 6: Personal Credit Scoring: Feature Importance
(a) Application: Average impact on model output (b) Application: Impact on model output
(c) Behavioral: Average impact on model output (d) Behavioral: Impact on model output
The combined network features also play an essential role in the final score; the feature sets
XEgoN et (D) and XGN N,N 2V (E) represent 25% and 33.4% in Application Scoring (See Figure 7(a)),
while the impact in Behavioral Scoring (See Figure 7(b)) are 18% and 28.3% respectively. In both
cases, the contribution of Node2Vec features is negligible. The network feature with the highest
impact is, as the average neighborhood’s amount, the network’s amount of unused revolving credit.
Personal Credit Scoring includes attributes generated with both FamilyNet and EOWNet net-
works. When analyzing the network-related features, almost all of them, in both scenarios, are
FamilyNet features. These results show us both the suitability of the network used to characterize
borrowers and the importance of the type of relationship used to build the network. In this study,
family ties are the most appropriate to characterize borrowers as regards the problem of individual
credit scoring.
22
Figure 7: Personal Credit Scoring: Treemap of Feature Importance, the average impact on model output
(a) Application Scoring
(b) Behavioral Scoring
7 Conclusions
This study presents an information processing methodology that allows us to assess the additional
value of social-interaction data to approach the credit scoring problem for thin-file clients. This
framework is applied in four scenarios arising from the consideration of all combinations of Appli-
cation and Behavioral Scoring of individual and business lending. Additionally, this methodology
allows the evaluation of different GRL approaches to feature extraction from social networks: hand-
crafted feature engineering, Node2Vec, and Graph Neural Networks. The results show an improve-
ment in creditworthiness assessment performance when different GRL approaches are combined.
Specifically, two of the three GRL methods significantly enhance credit scoring models, namely the
Hand-crafted Feature Engineering and Graph Neural Networks, which have the greatest impact
when used together. We believe this to be very relevant to the community because, until now,
these two methods have been used independently. On the other hand, we have found that the
contribution of Node2Vec is negligible. This result seems to justify the limited research conducted
on Node2Vec as a feature-engineering method for credit scoring.
As a baseline, we use a credit scoring model developed by a financial institution. This model,
the BenchScore, already outperforms the credit bureau model they obtain from the credit bureau
offices, and our methodology allows us to obtain better results in each of the four scenarios.
The highest value of the proposed approach is found in Unbanked Application Scoring. The
unbanked applicants, individuals, and companies, lack behavioral information, which, as it turns
out, is one of the best predictors of creditworthiness. Our approach overcomes the lack of behavioral
information and delivers a proper credit risk assessment using graph data. In this way, applicants
have greater access to the financial system. In the case of the Behavior Scoring models, our
methodology also improves performance. In both cases, the maximum improvement in predictive
performance is achieved when these GRL methods are used together.
23
Explanatory measures, such as SHAP values, allow us to understand each attribute’s contribu-
tion. If the impact on the output model is measured in this way, the baseline model (BenchScore),
although it continues to be an essential attribute, has a diminished effect because it is in the
presence of other good predictors. This feature importance analysis allows us to understand that
we cannot solely examine a company’s characteristics to evaluate the company, especially if it is
unbanked. We also have to understand that they are part of an ecosystem in which the owners,
suppliers, clients, and related companies are essential. The business ecosystem information allows
us to improve the creditworthiness assessment performance. A similar situation occurs in Personal
Credit Scoring, although with less intensity. The network data allows us to address the scarcity of
information and achieve a better credit risk assessment.
Our research shows that there is still room for improvement in incorporating network informa-
tion into the credit scoring problem. This methodology goes in the right direction, improving the
performance of creditworthiness assessment, and it has great value for unbanked and under-banked
people and even in the management of portfolio’s credit risk.
Acknowledgments
This work would not have been accomplished without the financial support of CONICYT-PFCHA
/ DOCTORADO BECAS CHILE / 2019-21190345. The second author acknowledges the support
of the Natural Sciences and Engineering Research Council of Canada (NSERC) [Discovery Grant
RGPIN-2020-07114]. This research was undertaken, in part, thanks to funding from the Canada
Research Chairs program. The last author thanks the partial support of FEDER funds through
the MINECO project TIN2017-85827-P and the European Union’s Horizon 2020 research and
innovation program under the Marie Sklodowska-Curie grant agreement No 777720.
References
Akoglu, H. (2018, 08). User’s guide to correlation coefficients. Turkish Journal of Emergency
Medicine, 18 (3), 91–93. doi: 10.1016/j.tjem.2018.08.001
Anderson, R. A. (2022). Credit intelligence and modelling: Many paths through the forest of credit
rating and scoring. Oxford University Press.
Arsov, N., & Mirceva, G. (2019). Network embedding: An overview. arXiv preprint
arXiv:1911.11726.
Aziz, S., & Dowling, M. (2019). Machine learning and ai for risk management. In Disrupting
finance: Fintech and strategy in the 21st century (pp. 33–50). Cham: Springer International
Publishing. doi: 10.1007/978-3-030-02330-0\ 3
Baidoo, E. (2020). A credit analysis of the unbanked and underbanked: an argument for alternative
data (PhD dissertation). Analytics and Data Science Institute, Kennesaw State University.
Bravo, C., Thomas, L. C., & Weber, R. (2015). Improving credit scoring by differentiating defaulter
behaviour. The Journal of the Operational Research Society, 66 (5), 771–781.
Bravo, C., & Óskarsdóttir, M. (2020). Evolution of credit risk using a personalized pagerank
algorithm for multilayer networks. arXiv preprint arXiv:2005.12418.
Breiman, L. (2001). Random forests. Machine learning, 45 (1), 5–32.
Carta, S., Ferreira, A., Reforgiato Recupero, D., & Saia, R. (2021). Credit scoring by leverag-
ing an ensemble stochastic criterion in a transformed feature space. Progress in Artificial
Intelligence, 10 (4), 417–432.
24
Cnudde, S. D., Moeyersoms, J., Stankova, M., Tobback, E., Javaly, V., & Martens, D. (2019).
What does your facebook profile reveal about your creditworthiness? using alternative data
for microfinance. Journal of the Operational Research Society, 70 (3), 353-363. doi: 10.1080/
01605682.2018.1434402
Cusmano, L. (2018). SME and entrepreneurship financing: The role of credit guarantee schemes
and mutual guarantee societies in supporting finance for small and medium-sized enterprises.
OECD SME and Entrepreneurship Papers, No. 1 . doi: 10.1787/35b8fece-en
Defferrard, M., Bresson, X., & Vandergheynst, P. (2016). Convolutional neural networks on graphs
with fast localized spectral filtering. In Advances in neural information processing systems
(pp. 3844–3852).
Diallo, B., & Al-Titi, O. (2017). Local growth and access to credit: Theory and evidence. Journal
of Macroeconomics, 54 , 410-423. (Banking in Macroeconomic Theory and Policy) doi:
https://doi.org/10.1016/j.jmacro.2017.07.005
Djeundje, V. B., Crook, J., Calabrese, R., & Hamid, M. (2021). Enhancing credit scoring with
alternative data. Expert Systems with Applications, 163 , 113766. doi: https://doi.org/
10.1016/j.eswa.2020.113766
Fang, F., & Chen, Y. (2019). A new approach for credit scoring by directly maximizing the
kolmogorov–smirnov statistic. Computational Statistics & Data Analysis, 133 , 180–194.
Fey, M., & Lenssen, J. E. (2019). Fast graph representation learning with pytorch geometric.
arXiv preprint arXiv:1903.02428.
Flach, P. A. (2012). Machine learning - the art and science of algorithms that make sense of data.
Cambridge University Press.
Freedman, S., & Jin, G. Z. (2017). The information value of online social networks: Lessons from
peer-to-peer lending. International Journal of Industrial Organization, 51 , 185 - 222. doi:
https://doi.org/10.1016/j.ijindorg.2016.09.002
Friedman, J. H. (2001). Greedy function approximation: a gradient boosting machine. Annals of
statistics, 1189–1232.
Goel, A., & Rastogi, S. (2021). Credit scoring of small and medium enterprises: a behavioural
approach. Journal of Entrepreneurship in Emerging Economies.
Grover, A., & Leskovec, J. (2016). Node2vec: Scalable feature learning for networks. In Proceedings
of the 22nd acm sigkdd international conference on knowledge discovery and data mining
(p. 855–864). New York, NY, USA: Association for Computing Machinery. doi: 10.1145/
2939672.2939754
Hagberg, A., Swart, P., & SChult, D. (2008). Exploring network structure, dynamics, and function
using networkx. In In proceedings of the 7th python in science conference (scipy (pp. 11–15).
Hamilton, W. L., Ying, R., & Leskovec, J. (2017). Representation learning on graphs: Methods
and applications. arXiv preprint arXiv:1709.05584.
Hurley, M., & Adebayo, J. (2017). Credit scoring in the era of big data. Yale Journal of Law and
Technology, 18 (1), 5.
Kaufman, S., Rosset, S., Perlich, C., & Stitelman, O. (2012, December). Leakage in data mining:
Formulation, detection, and avoidance. ACM Trans. Knowl. Discov. Data, 6 (4). doi: 10
.1145/2382577.2382579
Kipf, T. N., & Welling, M. (2016a). Semi-supervised classification with graph convolutional
networks. arXiv preprint arXiv:1609.02907.
Kipf, T. N., & Welling, M. (2016b). Variational graph auto-encoders. arXiv preprint
arXiv:1611.07308.
Kleinberg, J. M. (1999, 9). Authoritative sources in a hyperlinked environment. J. ACM , 46 (5),
25
604–632. doi: 10.1145/324133.324140
Kozodoi, N., Lessmann, S., Papakonstantinou, K., Gatsoulis, Y., & Baesens, B. (2019). A multi-
objective approach for profit-driven feature selection in credit scoring. Decision Support
Systems, 120 , 106-117. doi: https://doi.org/10.1016/j.dss.2019.03.011
Leskovec, J., & Sosič, R. (2016). Snap: A general-purpose network analysis and graph-mining
library. ACM Transactions on Intelligent Systems and Technology (TIST), 8 (1), 1.
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. In
Proceedings of the 31st international conference on neural information processing systems
(pp. 4765–4774). Red Hook, NY, USA: Curran Associates Inc.
Maldonado, S., Pérez, J., & Bravo, C. (2017). Cost-based feature selection for support vector
machines: An application in credit scoring. European Journal of Operational Research,
261 (2), 656 - 665. doi: https://doi.org/10.1016/j.ejor.2017.02.037
Moscato, V., Picariello, A., & Sperlı́, G. (2021). A benchmark of machine learning approaches
for credit score prediction. Expert Systems with Applications, 165 , 113986. doi: https://
doi.org/10.1016/j.eswa.2020.113986
Nargesian, F., Samulowitz, H., Khurana, U., Khalil, E. B., & Turaga, D. (2017). Learning feature
engineering for classification. In Proceedings of the twenty-sixth international joint conference
on artificial intelligence, IJCAI-17 (pp. 2529–2535). doi: 10.24963/ijcai.2017/352
Niu, B., Ren, J., & Li, X. (2019, 12). Credit scoring using machine learning by combing social
network information: Evidence from peer-to-peer lending. Information, 10 (12), 397. doi:
10.3390/info10120397
Óskarsdóttir, M., Bravo, C., Vanathien, J., & Baesens, B. (2018a, 11). Credit scoring for good:
enhancing financial inclusion with smartphone-based microlending. In Proceedings of the
thirty ninth international conference on information systems. San Francisco, California,
USA.
Óskarsdóttir, M., Bravo, C., Vanathien, J., & Baesens, B. (2018b, 7). Social network analytics in
micro-lending. In 29th european conference on operational research (08/07/18 - 11/07/18).
Valencia, Spain.
Óskarsdóttir, M., & Bravo, C. (2021). Multilayer network analysis for improved credit risk pre-
diction. Omega, 105 , 102520. doi: https://doi.org/10.1016/j.omega.2021.102520
Óskarsdóttir, M., Bravo, C., Sarraute, C., Vanthienen, J., & Baesens, B. (2019). The value
of big data for credit scoring: Enhancing financial inclusion using mobile phone data and
social network analytics. Applied Soft Computing, 74 , 26 - 39. doi: https://doi.org/10.1016/
j.asoc.2018.10.004
Óskarsdóttir, M., Bravo, C., Verbeke, W., Baesens, B., & Vanthienen, J. (2018). Effects of network
architecture on model performance when predicting churn in telco.
Óskarsdóttir, M., Bravo, C., Verbeke, W., Sarraute, C., Baesens, B., & Vanthienen, J. (2017).
Social network analytics for churn prediction in telco: Model building, evaluation and network
architecture. Expert Systems with Applications, 85 , 204 - 220. doi: https://doi.org/10.1016/
j.eswa.2017.05.028
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., . . . others (2019). Pytorch:
An imperative style, high-performance deep learning library. Advances in neural information
processing systems, 32 , 8026–8037.
Putra, S. G. P., Joshi, B., Redi, J., & Bozzon, A. (2020). A credit scoring model for smes based
on social media data. In M. Bielikova, T. Mikkonen, & C. Pautasso (Eds.), Web engineering
(pp. 113–129). Cham: Springer International Publishing.
Rabecca, H., Atmaja, N. D., & Safitri, S. (2018). Psychometric credit scoring in indonesia microfi-
26
nance industry: A case study in pt amartha mikro fintek. In The 3rd international conference
on management in emerging markets (icmem 2018) (pp. 620–631). Bali, Indonesia.
Rathi, S., Verma, J. P., Jain, R., Nayyar, A., & Thakur, N. (2022). Psychometric profiling of indi-
viduals using twitter profiles: A psychological natural language processing based approach.
Concurrency and Computation: Practice and Experience, e7029. doi: 10.1002/cpe.7029
Roa, L., Correa-Bahnsen, A., Suarez, G., Cortés-Tejada, F., Luque, M. A., & Bravo, C. (2021).
Super-app behavioral patterns in credit risk models: Financial, statistical and regulatory
implications. Expert Systems with Applications, 169 , 114486. doi: https://doi.org/10.1016/
j.eswa.2020.114486
Roa, L., Rodrı́guez-Rey, A., Correa-Bahnsen, A., & Valencia, C. (2021). Supporting financial
inclusion with graph machine learning and super-app alternative data.
Romero, D. M., Uzzi, B., & Kleinberg, J. (2019). Social networks under stress: Specialized team
roles and their communication structure. ACM Transactions on the Web (TWEB), 13 (1),
1–24.
Ruiz, S., Gomes, P., Rodrigues, L., & Gama, J. (2017). Credit scoring in microfinance using
non-traditional data. In E. Oliveira, J. Gama, Z. Vale, & H. Lopes Cardoso (Eds.), Progress
in artificial intelligence (pp. 447–458). Cham: Springer International Publishing.
Shumovskaia, V., Fedyanin, K., Sukharev, I., Berestnev, D., & Panov, M. (2020). Linking bank
clients using graph neural networks powered by rich transactional data.
Stevenson, M., Mues, C., & Bravo, C. (2021). The value of text for small business default
prediction: A deep learning approach. European Journal of Operational Research, 295 (2),
758-771. doi: https://doi.org/10.1016/j.ejor.2021.03.008
Sukharev, I., Shumovskaia, V., Fedyanin, K., Panov, M., & Berestnev, D. (2020). Ews-gcn: Edge
weight-shared graph convolutional network for transactional banking data.
Tan, T., & Phan, T. Q. (2018). Social media-driven credit scoring: The predictive value of social
structures. Available at SSRN 3217885 .
The Basel Committee on Banking Supervision. (2000, 09). Principles for the management of credit
risk. Basel Committee Publications, 75 .
The Global Financial Index. (2022). The global findex database 2021: Financial inclusion, digital
payments, and resilience in the age of covid-19. (Retrieved from https://openknowledge
.worldbank.org/bitstream/handle/10986/37578/9781464818974.pdf. Accessed July 3,
2022)
Thomas, L., Crook, J., & Edelman, D. (2017). Credit scoring and its applications. SIAM.
Vlasselaer, V. V., Bravo, C., Caelen, O., Eliassi-Rad, T., Akoglu, L., Snoeck, M., & Baesens,
B. (2015). Apate: A novel approach for automated credit card transaction fraud detection
using network-based extensions. Decision Support Systems, 75 , 38 - 48. doi: https://doi.org/
10.1016/j.dss.2015.04.013
Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2020). A comprehensive survey on
graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 1-21.
doi: 10.1109/TNNLS.2020.2978386
Zeng, G., & Zeng, E. (2019). On the three-way equivalence of auc in credit scoring with tied
scores. Communications in Statistics-Theory and Methods, 48 (7), 1635–1650.
Zhang, M., & Chen, Y. (2018). Link prediction based on graph neural networks. In Advances in
neural information processing systems (pp. 5165–5175).
27
A Feature Description
Table A1: Taxonomy of features used in the experiments and nomenclature that we used for the man-
agement of the data in the experiments.
Feature Set Nomenclature Prefix/suffix Description
Identifier N ODEAT T Feature subset identifier
Node AT T 01, · · · , AT T 04 The debt situation characterized by the
Features Borrower feature delinquency level
(XN ode ) AT T 05, · · · , AT T 08 The debt type: revolving, consumer,
identifier commercial, or mortgage
Other aspects of the customer’s debt,
AT T 09, · · · , AT T 13
payments in arrears, and the time in
the financial system
Bench Score The benchmark score
Identifier NodeStats Feature subset identifier
DegreeCentr Degree centrality
T riads Number of triads
Node P ageRank PageRank Algorithm
Statistics Statistic identifier
ArtP oint Articulation point
(XN odeStats ) Hits Auth Hits algorithm Authority score
Hits Hub Hits algorithm Hub score
EOWNET EOWNet Network
Network identifier
FamilyNet Family Network
Borrower feature AT T 01, · · · , AT T 13 Borrower Feature
identifier
EgoNetwork Network identifier
Agreggation
Aggregation MEAN Mean
Features
Function STD Standard Deviation
(XEgoN et )
Full All edges
Edges NotBridge Edges that are not bridges
IsBridge Edges that are bridges
Weighted W by + F eature Suffix for weighted aggregations
Aggregations
Identifier N2V Feature subset identifier
Node2Vec Embedding
Features EM B 01, · · · , EM B 08 The embedding number
Identifier
(XEgoN et ) EOWNET EOWNet Network
Network identifier
CHEB Graph Convolutional Network (GCN)
Graph Neural GNN Identifier
GAE Graph Autoencoder
Network Borrower feature 01, · · · , 13 Borrower Feature
Features identifier
(XGN N ) Embedding The embedding number, where n = 3
EM B 01, · · · , EM B n
Identifier for CHEB and n = 8 for GAE
Network identifier
FAMNET Family Network
28

On The Combination of Graph Data For Assessing Thin-File Borrowers' Creditworthiness

Uploaded by

Copyright:

Available Formats

On The Combination of Graph Data For Assessing Thin-File Borrowers' Creditworthiness

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

On The Combination of Graph Data For Assessing Thin-File Borrowers' Creditworthiness

Uploaded by

Copyright:

Available Formats

On the combination of graph data for assessing thin-file

Street, London, Ontario, N6A 3K7, Canada.

• We introduce a framework to combine traditional hand-engineered features with graph em-

2.2 Credit Risk and Social Networks

3 Representation Learning on Networks

3.1 Feature Engineering

3.2 Network Embeddings

3.3 Graph Neural Networks (GNN)

3.3.1 Graph Convolutional Networks

3.3.2 Derivation of GCN from Spectral Methods

Xi ∗ g = F −1 (F(Xi ) F(g)) (5)

where L̂ = 2L/λmax − I. Despite being a graph convolution simplification, this convolution

3.3.3 Graph Autoencoders (GAEs)

Â = σ(ZZ T ), with Z = GCN (X, A), (8)

4.1 Ethical and Privacy Protection Considerations

4.2 Social-Interaction Data

4.3 Financial Data

4.4 Network Construction

5 Experimental Design and Methodology

5.3 Traditional and Graph Representation Learning Features

• [EgoNet] EgoNetwork Agreggation: In this dataset, each node is characterized by the

Figure 1: Example of Network Features

train train train

sample sample sample sample

train model apply apply apply

sample sample sample sample

5.4 Feature Subsets

5.5 Evaluation Metrics

Another extensively utilized performance measure is the Kolmogorov-Smirnov (KS) statistic,

KS = max |PD (t) − PN D (t)| (9)

5.7 Experimental Setup

6 Results and Discussion

6.1 Implementation Details

6.2 Execution Time

Table 3: GCN and GAE Execution Time

6.3 Model Performance Results

6.3.1 Model Performance Using Traditional Features

6.3.2 Model Performance Using Graph Representation Learning Features

Table 6: Performance comparison of Business Application Scoring. Performance is measured by the

Table 7: Performance comparison of Business Application Scoring. Performance is measured by the

Table 9: Blended Graph Representation Learning performance. The performance enhancement of

6.4 Feature Importance Analysis

6.4.1 Business Credit Scoring

Figure 4: Business Credit Scoring: Feature Importance

(a) Application Scoring

(b) Behavioral Scoring

6.4.2 Personal Credit Scoring

Figure 6: Personal Credit Scoring: Feature Importance

(a) Application Scoring

(b) Behavioral Scoring

You might also like