DefaultGNN：一个用于从买卖方交易网络预测企业违约的双视角图神经网络框架

# DefaultGNN: A Dual-Perspective GNN Framework for Predicting Corporate Default from Buyer-Seller Transaction Networks
Source: [https://arxiv.org/html/2609.25542](https://arxiv.org/html/2609.25542)
Conference:Proceedings of the 35th ACM International Conference on Information and Knowledge Management; November 07–11, 2026; Rome, ItalyProceedings of the 35th ACM International Conference on Information and Knowledge Management \(CIKM ’26\), November 07–11, 2026, Rome, ItalyDOI:[10\.1145/3799682\.3840153](https://doi.org/10.1145/3799682.3840153)ISBN:979\-8\-4007\-2539\-5/2026/11CCS:Applied computing EconomicsCCS:Computing methodologies Artificial intelligence,Hyunsung KimAffiliation:KAIST,Daejeon,Republic of Koreaemail:[hyunsung\.kim@kaist\.ac\.kr](mailto:[email protected]),Seungyoon ChoiAffiliation:KAIST,Daejeon,Republic of Koreaemail:[csyoon08@kaist\.ac\.kr](mailto:[email protected]),KyoungYong ParkAffiliation:Techfin Ratings,Seoul,Republic of Koreaemail:[hess\_kpark@techfinratings\.com](mailto:[email protected]),Jihun LeeAffiliation:Techfin Ratings,Seoul,Republic of Koreaemail:[jihuny@techfinratings\.com](mailto:[email protected]),YongGu JiAffiliation:Douzone,Seoul,Republic of Koreaemail:[todcode@douzone\.com](mailto:[email protected])andChanyoung ParkNote:Corresponding author\.Affiliation:KAIST,Daejeon,Republic of Koreaemail:[cy\.park@kaist\.ac\.kr](mailto:[email protected])

© cc

###### Abstract\.

Corporate default prediction is a core problem in financial risk management, yet traditional credit models rely heavily on financial statements that are often sparse or unavailable for many firms\. Corporate transaction networks offer a complementary view of real economic activity, but how risk propagates through buyer–seller relationships remains underexplored\. We conduct a large\-scale empirical study using real\-world electronic tax\-invoice data spanning six years that links transaction histories with default events, revealing that transaction\-driven risk is both role\-dependent \(buyer or seller\) and scale\-dependent\. Based on these findings, we construct multiplex buyer\-view and seller\-view transaction networks and proposeDefaultGNN, a dual\-perspective graph neural network\-based framework for corporate default prediction\.DefaultGNNintegrates both views to model how risk flows through transactional relationships, achieving strong improvements over both attribute\-based and graph\-based baselines, especially for firms with limited intrinsic risk signals\. We further provide interpretable network\-based explanations by visualizing how distressed trading partners contribute to default risk\. In collaboration with a licensed credit rating agency, we validate thatDefaultGNN’s predictions complement existing credit scoring models, improving approval rates by 7–11%p without increasing default risk among approved firms\. The source code can be found at[https://github\.com/jhkim611/DefaultGNN](https://github.com/jhkim611/DefaultGNN)\.

###### Keywords:

Corporate Default Prediction, Graph Neural Networks, Transaction Networks, Financial Risk Propagation

††cc\-license:by## 1\.Introduction

Corporate credit evaluation\(ordefault prediction\), the task of determining whether a firm will default on its financial obligations in the future, plays a crucial role in financial decision\-making, including lending decisions and inter\-corporate trade activities\. Traditional corporate credit scoring models primarily rely on financial statements and historical default data\([Beaver, 1966](https://arxiv.org/html/2609.25542#bib.bib3);[Kumar et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib4)\), typically using statistical methods like logistic regression\([Altman, 1968](https://arxiv.org/html/2609.25542#bib.bib1);[Ohlson, 1980](https://arxiv.org/html/2609.25542#bib.bib2);[Hosmer Jr et al\., 2013](https://arxiv.org/html/2609.25542#bib.bib26);[Thomas et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib28);[Shumway, 2001](https://arxiv.org/html/2609.25542#bib.bib5)\)\. While these approaches are interpretable have shown stable performance, they are inherently constrained by the availability and timeliness of financial data\([Crook et al\., 2019](https://arxiv.org/html/2609.25542#bib.bib9);[Fuster et al\., 2022](https://arxiv.org/html/2609.25542#bib.bib10)\)\.

This limitation is particularly pronounced for small and medium enterprises \(SMEs\) and sole proprietors, who typically face infrequent financial reporting cycles and limited disclosure requirements\. As a result, financial statements for these firms are often missing, outdated or incomplete, leading to information asymmetry between financial institutions and the firms being evaluated\. In such settings, traditional financial statement\-driven credit scoring models tend to overestimate risk for SMEs111Such ”credit rationing” occurs when financial institutions impose limits on loan exposure due to difficulties in controlling risk through interest rate adjustments\.\([Stiglitz and Weiss, 1981](https://arxiv.org/html/2609.25542#bib.bib6);[Jaffee and Stiglitz, 1990](https://arxiv.org/html/2609.25542#bib.bib7);[Ghosh et al\., 2000](https://arxiv.org/html/2609.25542#bib.bib8)\), motivating growing interest in machine learning\-based models that incorporate non\-financial data in credit assessment systems\.

Among such non\-financial data, transactional relationships between businesses have garnered attention as they directly reflect the actual economic activity of firms\. Businesses do not operate in isolation; they are interconnected through buyer\-seller relationships\. The default of one partner can cascade through a transaction network, impacting the financial stability of adjacent firms\. Prior research has shown that credit risk contagion, i\.e\., the transfer of financial risk through transactions, can significantly influence default probability, with a meaningful correlation between a firm’s credit risk and that of its trading partners\([Giesecke and Weber, 2004](https://arxiv.org/html/2609.25542#bib.bib13);[Chen and He, 2012](https://arxiv.org/html/2609.25542#bib.bib14);[Berloco et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib15);[Jacobson and Von Schedvin, 2015](https://arxiv.org/html/2609.25542#bib.bib11);[Boissay and Gropp, 2007](https://arxiv.org/html/2609.25542#bib.bib12)\)\.

Despite this, large\-scale studies that leverage corporate transaction data for default prediction remain limited, as traditional statistical or linear models cannot effectively incorporate the relational and multi\-layered structure of interfirm transaction networks\. Recent advances in graph neural networks \(GNNs\)\([Kipf and Welling, 2016](https://arxiv.org/html/2609.25542#bib.bib16);[Veličković et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib17)\)have enabled relational modeling of financial systems, and have been applied to settings such as guarantee networks and supply chains for credit risk prediction\([Liu et al\., 2025](https://arxiv.org/html/2609.25542#bib.bib18);[Cheng et al\., 2020](https://arxiv.org/html/2609.25542#bib.bib20);[Wang et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib21)\)\. However, existing GNN\-based studies have largely underutilized transactional information\. In particular, most models either focus on non\-transactional relations \(e\.g\., loans or guarantees\) or simplify transactions and their accompanying risks to unidirectional and binary links, ignoring their bidirectional risk exposure and heterogeneous scale\.

Specifically, in real\-world transactions, risk would propagate in both directions:sellersface exposure to late or failed payments, whilebuyersare vulnerable to delivery failures that can disrupt downstream operations and other business relationships\. Moreover, larger transactions would likely induce higher risk than smaller ones\. Capturing these effects requires a more detailed analysis of how buyer–seller relationships and transaction magnitude shape default risk\.

To address this gap, we conduct a large\-scale empirical analysis of real\-world electronic tax\-invoice data, linking transaction histories with default events of trading partners \(see Section[3](https://arxiv.org/html/2609.25542#S3)\)\. We examine how default risk depends on who trades with whom, in which role \(buyer or seller\), and at what scale, and use these findings to construct multiplex buyer\- and seller\-view transaction networks\. These insights are then incorporated intoDefaultGNN, a dual\-perspective GNN framework for predicting corporate default from buyer\-seller transaction networks\.DefaultGNNintegrates both views, i\.e\., transactional patterns learned from separate GNN layers, to predict corporate default, enabling more accurate risk assessment especially for firms with limited intrinsic financial risk signals\.

Overall, our contributions can be summarized as follows:

- •Large\-scale empirical study of transaction\-driven default risk\.We analyze real\-world electronic tax\-invoice data spanning six years that link buyer\-seller transactions with default events, providing direct evidence for how risk propagates through corporate transaction networks\.
- •Role\- and scale\-aware modeling of transaction\-based risk\.We show that default risk depends on buyer/seller roles and transaction magnitude, and construct multiplex transaction networks that explicitly capture these nuances\.
- •Effectiveness\.Our proposedDefaultGNNintegrates buyer\-view and seller\-view embeddings to model how risk is transmitted through transactional relationships, enabling more accurate default prediction, especially for firms with limited financial data\.
- •Interpretable network\-based explanations of default risk\.DefaultGNNenables visualization of a firm’s transaction neighborhood, revealing how defaulted or distressed trading partners contribute to the predicted risk of the target firm, providing actionable insights for financial decision\-making\.
- •Practical validation with a credit rating agency\.In collaboration with Techfin Ratings, we validate thatDefaultGNN’s predictions complement an existing credit scoring model, improving approval rates without increasing default risk\.

## 2\.Related Works

### 2\.1\.Modeling for Corporate Default Prediction

Corporate default prediction has traditionally relied on financial statement\-based models, where bankruptcy risk is estimated using accounting ratios and firm\-level attributes\. Classical approaches such as the Altman Z\-score\([Altman, 1968](https://arxiv.org/html/2609.25542#bib.bib1)\)and the Ohlson O\-score\([Ohlson, 1980](https://arxiv.org/html/2609.25542#bib.bib2)\)remain widely used benchmarks due to their interpretability and strong theoretical grounding\. However, their applicability depends critically on the availability of detailed financial statements, which are often missing, delayed or unavailable for small and medium enterprises \(SMEs\) and private firms\([Stiglitz and Weiss, 1981](https://arxiv.org/html/2609.25542#bib.bib6);[Jaffee and Stiglitz, 1990](https://arxiv.org/html/2609.25542#bib.bib7);[Ghosh et al\., 2000](https://arxiv.org/html/2609.25542#bib.bib8)\)\.

To address these limitations, subsequent studies have explored machine learning\-based credit models, including logistic regression and tree\-based methods\([Hosmer Jr et al\., 2013](https://arxiv.org/html/2609.25542#bib.bib26);[Chen, 2016](https://arxiv.org/html/2609.25542#bib.bib27);[Thomas et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib28)\), which capture nonlinear relationships among firm attributes\. While these often improve predictive performance, they typically treat firms as independent entities and do not explicitly model inter\-firm relationships, overlooking the fact that default risk can propagate through economic connections\.

### 2\.2\.Network\-Based Financial Risk and GNNs

Recognizing that firms operate within interconnected economic systems, prior research has studied default risk contagion through inter\-firm networks, including trade relationships, supply chains and loan\-guarantee structures\. These studies provide empirical evidence that a firm’s default risk is correlated with that of its business partners, highlighting the importance of relational information\([Giesecke and Weber, 2004](https://arxiv.org/html/2609.25542#bib.bib13);[Chen and He, 2012](https://arxiv.org/html/2609.25542#bib.bib14);[Berloco et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib15);[Liu et al\., 2025](https://arxiv.org/html/2609.25542#bib.bib18);[Chen et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib19)\)\.

More recently, graph neural networks \(GNNs\)\([Kipf and Welling, 2016](https://arxiv.org/html/2609.25542#bib.bib16);[Veličković et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib17);[Wang et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib21)\)have enabled scalable learning on relational financial data and have been applied to settings such as loan\-guarantee networks and supply\-chain risk assessment, often outperforming traditional attribute\-based models\([Xu et al\., 2022](https://arxiv.org/html/2609.25542#bib.bib22);[Bi et al\., 2022](https://arxiv.org/html/2609.25542#bib.bib23);[Sun, 2024](https://arxiv.org/html/2609.25542#bib.bib24);[Yang et al\., 2021](https://arxiv.org/html/2609.25542#bib.bib25)\)\. Notably, DGANN\([Cheng et al\., 2020](https://arxiv.org/html/2609.25542#bib.bib20)\)leverages guarantee relationships to model default risk in lending systems\.

Despite these advances, existing network analyses and GNN\-based approaches often focus on non\-transactional relations or represent transactions as single\-view, unweighted and unidirectional graphs, overlooking key characteristics of real\-world corporate transactions\. In particular, buyer\-seller perspectives and transaction magnitude, which are critical to understanding how risk propagates through transaction networks, remain largely underexplored\. Our work addresses this gap by combining large\-scale empirical analysis of transaction data with role\- and scale\-aware network modeling\.

## 3\.Data Description and Analysis

### 3\.1\.Data Collection and Preprocessing

The datasets used in this study consist of firm\-level attributes, inter\-firm transaction records, and corporate default information spanning six years \(2018–2023\)\. Firm attributes and transaction data were obtained from a large\-scale electronic tax\-invoice system operated by Douzone Bizon, which is widely used by businesses in Korea for statutory reporting\. Corporate default labels were integrated based on credit event records compiled by a licensed credit rating agency in accordance with Basel II default definitions\. Preprocessing steps include removing firms with invalid identifiers or missing business type, discarding non\-positive transaction amounts, and aggregating multiple transactions between the same pair of firms within each month by transaction role, retaining at most two directed interactions per firm pair \(each firm acting as the seller\) with transaction amounts summed accordingly\. Data collection and usage were conducted within a secure, access\-controlled environment provided by the data owner\.

### 3\.2\.Default\-Transaction Correlation

#### 3\.2\.1\.Effect of Partner Default

We first examine how the default history of trading partners relates to a firm’s future default risk\. At the transaction level, transactions involving partners with prior \(i\.e\., in the previous 12 months\) default history are consistently associated with a higher likelihood of future \(i\.e\., in the next 12 months\) default of the target firm across all years \(see Fig\.[1](https://arxiv.org/html/2609.25542#S3.F1)\)\.

This pattern persists when aggregating transactions at the firm level\. As shown in Fig\.[2](https://arxiv.org/html/2609.25542#S3.F2), firms that transact with at least one defaulted partner exhibit substantially higher future default rates than those whose partners have no default history\. Moreover, firms that eventually default tend to have both a higher number and a higher proportion of defaulted trading partners among their relationships \(see Fig\.[3](https://arxiv.org/html/2609.25542#S3.F3)\)\. These trends remain consistent across time\.

Figure 1\.Transactions involving partners with default history are more likely to cause future default of the target firm\.Figure 2\.Firms with defaulted trading partners exhibit higher future default rates than those whose partners have no prior default history\.Overall, these results indicate astrong correlation between partner default history and future corporate default, underscoring the importance of modeling inter\-firm dependencies\.

Figure 3\.Firms that later default tend to have more defaulted trading partners\.
#### 3\.2\.2\.Effect of Transaction Role

A transaction can be viewed from two complementary perspectives: as a sale for the seller and as a purchase for the buyer\. While prior work often models default propagation in a single direction — typically emphasizing payment risk faced by sellers, analogous to loaners exposed to guarantee risk\([Cheng et al\., 2020](https://arxiv.org/html/2609.25542#bib.bib20)\)— such a unidirectional view overlooks risk exposure on the buyer side, where delivery failures or operational disruptions can also have cascading effects\.

Empirical results support the relevance of both perspectives\. As shown in Fig\.[4](https://arxiv.org/html/2609.25542#S3.F4), both sales to and purchases from defaulted partners are associated with elevated future default risk\. This suggests thatdefault risk propagates through transactions in a bidirectional manner, motivating the need to explicitly distinguish buyer\-view and seller\-view transaction networks\.

Figure 4\.Bothsales to and purchases from defaulted partners are associated with increased future default risk, indicating bidirectional risk propagation\.Figure 5\.Large\-scale \(proportion≥0\.5\\geq 0\.5\) transactions with defaulted partners contribute substantially more to future default risk than small\-scale \(proportion<0\.5<0\.5\) transactions\.
#### 3\.2\.3\.Effect of Transaction Scale

We further investigate how transaction magnitude influences default propagation\. Transaction scale is defined as the ratio of a transaction’s amount to the firm’s total transaction volume within the same month\.

As shown in Fig\.[5](https://arxiv.org/html/2609.25542#S3.F5), firms that engage in large\-scale transactions with defaulted partners are significantly more likely to experience future default than firms without such transactions\. In contrast, small\-scale transactions with defaulted partners exhibit a much weaker effect\. This demonstrates thattransaction magnitude plays a critical role in shaping default riskand should be explicitly incorporated into relational modeling\.

#### 3\.2\.4\.Summary of Analysis

In summary, our empirical analysis identifies three key factors underlying default propagation in transaction networks: 1\) the default history of trading partners, 2\) the transactional role of firms as buyers or sellers, and 3\) the relative scale of transactions\. These results indicate that default risk propagates through transactions in a bidirectional and heterogeneous manner, with economically significant transactions contributing disproportionately to future default risk\. All observed trends remain consistent across the years considered\.

These findings highlight the limitations of unidirectional or binary relational modeling and motivate a dual\-perspective, edge\-weighted graph\-based approach that explicitly accounts for heterogeneous risk exposure\.

### 3\.3\.Multiplex Transaction Networks

Based on the transactional data, we construct multiplex inter\-firm transaction networks that explicitly distinguish buyer–seller roles and transaction scale\. For each year, firms are represented as nodes, and transactions between firms are represented as directed edges\. To reflect asymmetric risk propagation mechanisms, we define two complementary transaction views over the same set of firms\.

In theseller\-viewtransaction network, a directed edge from firmuuto firmvvrepresents a transaction wherevvacts as the seller anduuacts as the buyer\. This view captures risk exposure arising from delayed or failed payments, where financial stress at the buyer can propagate upstream to the seller\. Conversely, in thebuyer\-viewtransaction network, a directed edge fromvvtouurepresents the same transaction, reflecting the buyer’s exposure to delivery failures or operational disruptions at the seller\.

To account for heterogeneous economic impact, each edge is assigned a weight reflecting the relative transaction scale\. Specifically, for a transaction between firmsuuandvv, the edge weight is defined as the ratio of the transaction amount to the total sales \(seller view\) or total purchases \(buyer view\) of the corresponding firm within the same period, i\.e\., month\. To mitigate the effect of extreme values, the weights can be further scaled by the following function:wu​vn​e​w=l​o​g​\(1\+α⋅wu​vo​r​i​g\)/l​o​g​\(1\+α\)w\_\{uv\}^\{new\}=log\(1\+\\alpha\\cdot w\_\{uv\}^\{orig\}\)/log\(1\+\\alpha\), wherewu​vo​r​i​gw\_\{uv\}^\{orig\}denotes the initial relative transaction ratio andα\\alphais a scaling factor\.

The resulting multiplex representation consists of two directed, edge\-weighted graphs sharing the same node set but encoding distinct transactional semantics \(see Table[1](https://arxiv.org/html/2609.25542#S5.T1)for summary statistics\)\. By jointly modeling these buyer\- and seller\-view transaction networks, this representation captures role\- and scale\-dependent risk propagation patterns inherent in real\-world corporate transactions\.

### 3\.4\.Firm Attributes

Each firm is represented by a set of non\-financial firm attributes \(108\-dimensional\) that capture basic operational characteristics under realistic data constraints where detailed financial statements are unavailable\. These include a recent \(within that year\) default indicator, taxation category, business type encodings at multiple levels of granularity, and log\-scaled aggregate sales and purchase statistics computed from historical transaction records\. Importantly, these features do not encode information about specific trading partners or network structure, serving as a baseline that is substantially enhanced by the relational information from transaction networks\.

## 4\.Proposed Framework:DefaultGNN

### 4\.1\.Problem Setting and Overview

Let𝒟=\(𝒱,ℰ\)\\mathcal\{D=\(V,E\)\}be a yearly corporate transaction dataset\.𝒱\\mathcal\{V\}is a set of nodes, each node corresponding to an individual firm\.xv∈ℝFx\_\{v\}\\in\\mathbb\{R\}^\{F\}denotes the node features of each nodev∈𝒱v\\in\\mathcal\{V\}, consisting ofFFfirm attributes\. Each node is further labeled asyv∈\{0,1\}y\_\{v\}\\in\\\{0,1\\\}based on whether the corresponding firm defaults within the following year, i\.e\.,yv=1y\_\{v\}=1for defaulted firms/nodes andyv=0y\_\{v\}=0for non\-defaulted firms\.ℰ=\[Es​e​l​l,Eb​u​y\]\\mathcal\{E\}=\[E^\{sell\},E^\{buy\}\]contains two edge sets, corresponding to the seller\-view and buyer\-view transaction networks defined in Section[3\.3](https://arxiv.org/html/2609.25542#S3.SS3), respectively\. Specifically, each directed edge\(u,v\)\(u,v\)inEv​i​e​wE^\{view\}\(v​i​e​w∈\{s​e​l​l,b​u​y\}view\\in\\\{sell,buy\\\}\) is paired with edge weightwu​vw\_\{uv\}, where the directions and weights in each set are defined to represent their respective view\. In other words,𝒟\\mathcal\{D\}contains year\-wise multiplex transaction networks for a set of firms\. For clarity, we additionally denote𝒢s​e​l​l=\(𝒱,Es​e​l​l\)\\mathcal\{G\}^\{sell\}=\(\\mathcal\{V\},E^\{sell\}\)and𝒢b​u​y=\(𝒱,Eb​u​y\)\\mathcal\{G\}^\{buy\}=\(\\mathcal\{V\},E^\{buy\}\)as the seller\- and buyer\-view graphs, respectively, where𝒱\\mathcal\{V\}is shared among the two graphs\.

Given dataset𝒟=\(𝒱,ℰ\)\\mathcal\{D=\(V,E\)\}, our objective is to learn a binary classifier for corporate default prediction\. In detail, the classifier is first trained on a training subset𝒱t​r​a​i​n\\mathcal\{V\}\_\{train\}and validated on a validation subset𝒱v​a​l\\mathcal\{V\}\_\{val\}, finally evaluated by predicting the node labels in a test subset𝒱t​e​s​t\\mathcal\{V\}\_\{test\}, whose labels were not available during training\.

Our proposedDefaultGNNis a dual\-perspective GNN\-based framework \(see Fig\.[6](https://arxiv.org/html/2609.25542#S4.F6)\) that\(1\)learns view\-specific representations from each transaction network,\(2\)integrates them via a role\-adaptive gating mechanism further stabilized through a view\-consistency regularizer, and finally\(3\)utilizes the fused embeddings to identify future defaults of nodes\.

![Refer to caption](https://arxiv.org/html/2609.25542v1/architecture.png)Figure 6\.Overall framework ofDefaultGNN\. View\-specific embeddings learned from multiplex transaction networks are fused via gating, regularized through view consistency\. The fused embeddings are then used for default prediction\.
### 4\.2\.View\-Specific Graph Encoders

DefaultGNNemploys two parallel graph encoders, one for each transaction view\. Both encoders share the same architectural design but operate on different graph structures, and map firms and their relationships into latent representations that capture default risk propagation patterns specific to the corresponding transaction role\.

Specifically, each encoder consists of multiple GNN layers that perform structured message passing over the transaction graph, aggregating information from trading partners while accounting for transaction scale represented through edge weights\. Formally, node representation for a firm nodevvat layerllis updated as:

\(1\)hv\(l\+1\)=ϕ⁡\(∑u∈𝒩⁡\(v\)∪\(v\)αu​v​\(wu​v\)​𝐖\(l\)​hu\(l\)\),h\_\{v\}^\{\(l\+1\)\}=\\phi\\left\(\\sum\_\{u\\in\\mathcal\{N\}\(v\)\\cup\(v\)\}\\alpha\_\{uv\}\(w\_\{uv\}\)\\mathbf\{W\}^\{\(l\)\}h\_\{u\}^\{\(l\)\}\\right\),wherehu\(l\)h\_\{u\}^\{\(l\)\}denotes the representation of nodeuu\(hu\(0\)=xuh\_\{u\}^\{\(0\)\}=x\_\{u\}initially\),𝒩⁡\(v\)\\mathcal\{N\}\(v\)is the set of 1\-hop neighbor nodes of nodevv,𝐖\(l\)∈ℝd×d\\mathbf\{W\}^\{\(l\)\}\\in\\mathbb\{R\}^\{d\\times d\}is the learnable weight matrix \(ddis the latent embedding dimension size and𝐖\(0\)∈ℝd×F\\mathbf\{W\}^\{\(0\)\}\\in\\mathbb\{R\}^\{d\\times F\}\),wu​vw\_\{uv\}is the edge weight corresponding to directed edge\(u,v\)\(u,v\),αu​v​\(wu​v\)\\alpha\_\{uv\}\(w\_\{uv\}\)is a normalized aggregation coefficient that depends on both graph structure and transaction scale, andϕ⁡\(⋅\)\\phi\(\\cdot\)is a non\-linear activation\. By incorporating transaction magnitude directly into message passing, the encoders capture heterogeneous risk exposure arising from economically significant trading relationships, consistent with our empirical analysis\. After the final GNN layer, the resulting representations are taken as the view\-specific embeddings:hvs​e​l​lh\_\{v\}^\{sell\}andhvb​u​yh\_\{v\}^\{buy\}\.

### 4\.3\.Role\-Adaptive Gated Fusion

Given the seller\- and buyer\-view embeddingshvs​e​l​lh\_\{v\}^\{sell\}andhvb​u​yh\_\{v\}^\{buy\},DefaultGNNintegrates them using a gating mechanism that adaptively determines their relative importance on a per\-firm basis\. The fusion gate is computed as \(σ⁡\(⋅\)\\sigma\(\\cdot\)is the sigmoid function\):

\(2\)gv=σ\(wg​a​t​ev⊺\[hvs​e​l​l\|\|hvb​u​y\]\+b\)\.g\_\{v\}=\\sigma\(w\_\{gate\_\{v\}\}^\{\\intercal\}\[h\_\{v\}^\{sell\}\|\|h\_\{v\}^\{buy\}\]\+b\)\.The gated embeddings are computed ash~vs​e​l​l=gv⋅hvs​e​l​l\\tilde\{h\}\_\{v\}^\{sell\}=g\_\{v\}\\cdot h\_\{v\}^\{sell\}andh~vb​u​y=\(1−gv\)⋅hvb​u​y\\tilde\{h\}\_\{v\}^\{buy\}=\(1\-g\_\{v\}\)\\cdot h\_\{v\}^\{buy\}, respectively, and concatenated to form the fused representationzv=\[h~vs​e​l​l\|\|h~vb​u​y\]z\_\{v\}=\[\\tilde\{h\}\_\{v\}^\{sell\}\|\|\\tilde\{h\}\_\{v\}^\{buy\}\]\.zvz\_\{v\}is then passed to a linear classifier to predict default probability\. This gating mechanism allowsDefaultGNNto account for heterogeneity in transactional roles, enabling the model to emphasize buyer\-side or seller\-side risk exposure depending on the firm’s position in the transaction network\.

View\-Consistency Regularization\.To promote stable integration of complementary transaction views, we introduce aview\-consistency regularizerthat encourages compatible risk estimates when both views are confident, guiding the fusion process toward coherent representations\.

Specifically,DefaultGNNadditionally produces view\-specific default probabilities:pvs​e​l​l=P⁡\(y=1\|hvs​e​l​l\),pvb​u​y=P⁡\(y=1\|hvb​u​y\)p\_\{v\}^\{sell\}=P\(y=1\|h\_\{v\}^\{sell\}\),p\_\{v\}^\{buy\}=P\(y=1\|h\_\{v\}^\{buy\}\), from which we can define view\-specific confidences in range\[0,1\]\[0,1\]:cvs​e​l​l=2​\|pvs​e​l​l−0\.5\|,cvb​u​y=2​\|pvb​u​y−0\.5\|c\_\{v\}^\{sell\}=2\|p\_\{v\}^\{sell\}\-0\.5\|,c\_\{v\}^\{buy\}=2\|p\_\{v\}^\{buy\}\-0\.5\|\. Disagreement between the obtained probabilities is penalized through a confidence\-weighted mean squared error:

\(3\)ℒc​o​n​s=∑vwvc​\(pvs​e​l​l−pvb​u​y\)2∑vwvc,wvc=cvs​e​l​l​cvb​u​y\.\\mathcal\{L\}\_\{cons\}=\\frac\{\\sum\_\{v\}w\_\{v\}^\{c\}\(p\_\{v\}^\{sell\}\-p\_\{v\}^\{buy\}\)^\{2\}\}\{\\sum\_\{v\}w\_\{v\}^\{c\}\},w\_\{v\}^\{c\}=c\_\{v\}^\{sell\}c\_\{v\}^\{buy\}\.
This regularization encourages agreement primarily when both views are confident, while allowing divergence when either view is uncertain\. As a result, it stabilizes gated fusion without suppressing view\-specific representations or forcing collapse\.

### 4\.4\.Training Objective

DefaultGNNis trained using weighted binary cross entropy loss to address the severe class imbalance inherent in our task, with far less firms experiencing future default \(see Table[1](https://arxiv.org/html/2609.25542#S5.T1)\)\. Formally, given prediction outputsp^v=σ⁡\(zv\)\\hat\{p\}\_\{v\}=\\sigma\(z\_\{v\}\)the classification loss is defined as:

\(4\)ℒc​l​a​s​s=−∑v∈𝒱t​r​a​i​n\(wb​c​e⋅yv⋅log\(p^v\)\+\(1−wb​c​e\)⋅\(1−yv\)⋅log\(1−p^v\),\\mathcal\{L\}\_\{class\}=\-\\sum\_\{v\\in\\mathcal\{V\}\_\{train\}\}\(w\_\{bce\}\\cdot y\_\{v\}\\cdot log\(\\hat\{p\}\_\{v\}\)\+\(1\-w\_\{bce\}\)\\cdot\(1\-y\_\{v\}\)\\cdot log\(1\-\\hat\{p\}\_\{v\}\),wherewb​c​e∈\[0,1\]w\_\{bce\}\\in\[0,1\]is assigned to give larger emphasis to correctly identifying defaulted firms\.

The overall training objective isℒ=ℒc​l​a​s​s\+λ​ℒc​o​n​s\\mathcal\{L\}=\\mathcal\{L\}\_\{class\}\+\\lambda\\mathcal\{L\}\_\{cons\}, whereλ\\lambdais a parameter controlling the strength of regularization\. Through this procedure,DefaultGNNeffectively integrates both buyer\- and seller\-view transaction networks to predict future corporate default\.

## 5\.Experiments

In this section, we conduct comprehensive experiments to answer the following research questions:

- •RQ1\.How well does our proposedDefaultGNNperform in predicting corporate default compared with baselines?
- •RQ2\.How effective are our multiplex networks and additional modules in enhancingDefaultGNN’s performance?
- •RQ3\.IsDefaultGNNrobust under out\-of\-time settings?
- •RQ4\.CanDefaultGNNidentify and visualize distressed trading partners that critically contribute to a target firm’s default?

Table 1\.Dataset statistics\.### 5\.1\.Experimental Setup

Datasets\.We evaluateDefaultGNNon our constructed multiplex transaction networks \(see Section[3\.3](https://arxiv.org/html/2609.25542#S3.SS3)for details\) spanning six years, i\.e\., six distinct datasets\. The statistics can be found in Table[1](https://arxiv.org/html/2609.25542#S5.T1)\. Note that the statistics are equivalent for both buyer\- and seller\-view networks\. 70%, 10% and 20% of the firms are assigned to the training, validation and test sets, respectively, where the ratio of defaulted firms is kept consistent among them, e\.g\., for year 2021 they each contain 5,451, 778 and 1,559 defaulted nodes \(all 0\.79% of their corresponding sets\), respectively\.

Baselines\.We compare against baselines spanning four categories\.Attribute\-based methods \(G1\)\(Logistic Regression\([Hosmer Jr et al\., 2013](https://arxiv.org/html/2609.25542#bib.bib26)\)and XGBoost\([Chen, 2016](https://arxiv.org/html/2609.25542#bib.bib27)\)\) treat firms as independent instances, only utilizing firm attributes\. We also include variants that incorporate additional transaction\-related information as features \("\+ Trans\. Rel\.": transaction counts with partners grouped by business type, and "\+ Def\. Agg\.": monthly and annual counts of defaulted partners\)\.Standard GNNs \(G2\)\(GCN\([Kipf and Welling, 2016](https://arxiv.org/html/2609.25542#bib.bib16)\), GAT\([Veličković et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib17)\)and DGANN\([Cheng et al\., 2020](https://arxiv.org/html/2609.25542#bib.bib20)\): a directed graph attention network predicting defaults on loan\-guarantee networks\) model firms as nodes and transactions as edges, and are applied on the seller\-view graph\.Multi\-relational GNNs \(G3\)\(R\-GCN\([Schlichtkrull et al\., 2018](https://arxiv.org/html/2609.25542#bib.bib37)\)and CompGCN\([Vashishth et al\., 2019](https://arxiv.org/html/2609.25542#bib.bib38)\)\) handle multiple edge types within a single model, operating on a combined graph where buyer and seller transactions are treated as distinct relation types\.Multiplex graph models \(G4\)\(DMGI\([Park et al\., 2020](https://arxiv.org/html/2609.25542#bib.bib39)\), DMG\([Mo et al\., 2023](https://arxiv.org/html/2609.25542#bib.bib40)\)and MGHC\([Huang et al\., 2025](https://arxiv.org/html/2609.25542#bib.bib41)\), adapted to the supervised setting\) maintain separate encoders for each view\-specific graph, mirroringDefaultGNN’s dual\-encoder design\.

Implementation Details\.DefaultGNNis implemented in PyTorch 2\.2\.1 with CUDA 12\.1 and trained on an NVIDIA RTX A6000 GPU\. The view\-specific graph encoders use a 2\-layer GCN\([Kipf and Welling, 2016](https://arxiv.org/html/2609.25542#bib.bib16)\)with embedding dimensiond=128d=128\. The edge scaling factorα\\alpha, consistency regularization weightλ\\lambda, and BCE weightwb​c​ew\_\{bce\}are set to 50, 0\.05, and 0\.9, respectively\. All parameters are optimized with Adam\([Kingma, 2014](https://arxiv.org/html/2609.25542#bib.bib36)\)\(learning rate 0\.001\) for up to 500 epochs with early stopping \(patience 10\)\. For all baselines, we follow the original implementations and suggested hyperparameters\.

Table 2\.Overall model performance\. The best AR score for each dataset is highlighted in bold\. Standard deviations forDefaultGNN:±\\pm0\.001–0\.003 \(All\),±\\pm0\.001–0\.005 \(NoHist\), with comparable ranges for other methods\.DefaultGNNsignificantly outperforms the best baseline on all datasets \(Wilcoxon signed\-rank test,p=0\.016p=0\.016\)\.Evaluation Details\.We evaluate model performance using the Accuracy Ratio \(AR\) metric derived from the Area Under the Curve \(AUC\) score, formally defined asA​R=2×A​U​C−1AR=2\\times AUC\-1\(range: \[0, 1\]\)\. AR is a standard performance metric in credit risk modeling, widely used in industry\([Engelmann et al\., 2003](https://arxiv.org/html/2609.25542#bib.bib29);[Basel Committee on Banking Supervision, 2005](https://arxiv.org/html/2609.25542#bib.bib30);[Thomas et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib28)\)to assess the discriminatory power of default prediction models, especially under severe class imbalance where defaults are rare\.

In practical credit assessment, many firms — especially SMEs and private firms — have no prior default history, limiting the effectiveness of models that rely on historical risk signals\. To evaluate performance in such data\-poor settings, we report the AR score not only on all test firms \(All\), but also on the subset of firms with no prior default history \(NoHist\)\. Performance on the latter reflects a model’s ability to infer emerging risk from relational transaction patterns rather than intrinsic historical indicators\. For all experiments, we report the average performance of 3 independent runs\.

### 5\.2\.Performance Comparison \(RQ1\)

Our main results can be found in Table[2](https://arxiv.org/html/2609.25542#S5.T2)\. Attribute\-based methods \(G1\) consistently underperform graph\-based approaches by a large margin, even when augmented with transaction\-related features, highlighting that encoding transactional data as simple features cannot capture the complex relational structure between firms\.

Among graph\-based methods, standard GNNs \(G2\) on the seller\-view graph achieve strong results, demonstrating the value of relational modeling\. However, multi\-relational GNNs \(G3\) perform comparably to or even below standard GNNs, suggesting that merging buyer–seller signals within each GNN layer loses the view\-specific information needed to capture asymmetric risk propagation\.

Multiplex graph models \(G4\) consistently outperform both G2 and G3, confirming the importance of preserving view\-specific representations\. Nevertheless,DefaultGNNoutperforms all G4 baselines across every year and both evaluation settings, despite sharing the same dual\-encoder architecture\. We attribute this to the domain\-aligned gated fusion:DefaultGNN’s learned gate values exhibit a strong positive correlation \(Pearsonr=0\.73,p<10−6r=0\.73,p<10^\{\-6\}\) with firms’ transaction role balance, measured by annual sales / \(annual sales\+\+purchases\)\. This indicates that the model adaptively emphasizes the seller\- or buyer\-view embedding depending on each firm’s economic role, assigning higher weights to the seller view for seller\-dominant firms and vice versa\. In contrast, more generic fusion strategies such as averaging \(DMGI\) or disentanglement \(DMG\) do not capture this role\-dependent structure\.

DefaultGNN’s advantage is particularly pronounced for firms without prior default history \(NoHist\), the most practically important setting where traditional credit models fail\. This reinforces that DefaultGNN effectively captures emerging risk from transactional relationships even in the absence of intrinsic historical risk signals\.

Table 3\.Ablation studies onDefaultGNN\(AR score\)\.
### 5\.3\.Ablation Study \(RQ2\)

To assess the contribution of each component inDefaultGNN, we conduct ablation experiments on the 2021 dataset \(see Table[3](https://arxiv.org/html/2609.25542#S5.T3)\), with consistent trends observed across other years\.Network variantsisolate individual graph\-related design choices: bothDefaultGNN\+SellerView andDefaultGNN\+BuyerView use only a single transaction view;DefaultGNN\+CollapsedView1 combines edges from both views into a single directed graph while forDefaultGNN\+CollapsedView2 each transaction is reduced to a single undirected edge with both view weights as edge attributes;DefaultGNN−\-EdgeWeights removes edge weights entirely\.Fusion variantsmodify the integration mechanism:DefaultGNN−\-Gating replaces gated fusion with concatenation,DefaultGNN−\-ConsReg removes the consistency regularizer, andDefaultGNN\+CrossAttn replaces gating with bidirectional cross\-attention between views\.

The fullDefaultGNNachieves the highest performance over all variants\. Both single\-view models deteriorate similarly, indicating that both perspectives are equally important\. Both collapsed views result in a larger drop, highlighting the necessity of separately modeling buyer and seller views\. Removing edge weights degrades performance below even single\-view models, confirming that transaction scale plays a critical role in default prediction\.

Among the fusion variants, replacing gating with concatenation or removing the regularizer both degrade performance while still outperforming single\-view models\. Replacing gating with cross\-attention causes an even larger drop, confirming that preserving view\-specific representations and adjusting only their relative importance is more effective than allowing views to modify each other\.

Further,DefaultGNNis robust across hyperparameter choices: varying embedding dimensiond∈\{32,64,d\\in\\\{32,64,128,256\}128,256\\\}, regularization weightλ∈\{0\.01,0\.05,0\.1,0\.2\}\\lambda\\in\\\{0\.01,0\.05,0\.1,0\.2\\\}, and edge scaling factorα∈\{1,10,50,100,200\}\\alpha\\in\\\{1,10,50,100,200\\\}, the NoHist AR ranges from 0\.452 to 0\.472 on the 2021 dataset, consistently outperforming the best baselines\.

\(a\)Figure 7\.Comparison of model performances under OOT settings\. Bar heights represent each model’s AR score on the NoHist subset, while values in the bars show how much the OOT performance of each model diverges from the original performance, i\.e\., training and testing on the same year\.
### 5\.4\.Out\-of\-time Analysis \(RQ3\)

To evaluate generalization in a realistic credit risk assessment setting where models are trained on historical data and tested on future years\([Shumway, 2001](https://arxiv.org/html/2609.25542#bib.bib5);[Hand and Henley, 1997](https://arxiv.org/html/2609.25542#bib.bib31);[Baesens et al\., 2003](https://arxiv.org/html/2609.25542#bib.bib32);[Engelmann et al\., 2003](https://arxiv.org/html/2609.25542#bib.bib29)\), we compareDefaultGNNagainst GCN and XGBoost in two out\-of\-time \(OOT\) configurations: \(1\) training on a single year \(2018\) and testing on 2019–2023, and \(2\) training sequentially on three years \(2018–2020\) and testing on 2021–2023\.

As shown in Fig\.[7](https://arxiv.org/html/2609.25542#S5.F7),DefaultGNNconsistently outperforms baselines across both configurations\. In the single\-year setting,DefaultGNNmaintains strong performance even as the training\-test gap increases to five years, whereas GCN and XGBoost exhibit substantially larger degradation\. In the three\-year setting, sequential retraining allowsDefaultGNNto closely match or even improve upon its in\-year performance, demonstrating effective integration of new data over time\. These results confirm thatDefaultGNNgeneralizes well to future unseen data, a critical requirement for real\-world credit risk models that must adapt to evolving economic conditions\.

### 5\.5\.Case Studies \(RQ4\)

We present two case studies in Fig\.[8](https://arxiv.org/html/2609.25542#S5.F8)to illustrate howDefaultGNNidentifies meaningful transaction\-based risk signals beyond raw prediction performance\. Specifically, for each target firm, we extract a two\-hop local transaction subgraph from both buyer and seller views and attribute the prediction to individual transaction edges using integrated gradients\([Sundararajan et al\., 2017](https://arxiv.org/html/2609.25542#bib.bib33);[Ying et al\., 2019](https://arxiv.org/html/2609.25542#bib.bib34);[Pope et al\., 2019](https://arxiv.org/html/2609.25542#bib.bib35)\), which identifies the trading relationships and partners that most strongly drive the model’s decision\.

In Fig\.[8](https://arxiv.org/html/2609.25542#S5.F8)\(a\),DefaultGNNhighlights a transaction with a previously defaulted firm located two hops away from the target firm, demonstrating its ability to capture indirect risk propagation\. In contrast, XGBoost fails to incorporate this dependency and incorrectly predicts the firm as safe\. In Fig\.[8](https://arxiv.org/html/2609.25542#S5.F8)\(b\), an influential relationship corresponds to a buyer\-view transaction with a defaulted trading partner, which is then propagated to the target firm through a seller\-view transaction\. WhileDefaultGNNcorrectly captures this signal through its dual\-view design, a single\-view GCN misses the risk and produces an incorrect prediction, underscoring the importance of modeling asymmetric transaction roles\.

Beyond improving predictive accuracy, these visualizations offer clear and actionable insights by identifying influential trading partners and transaction links\. Such interpretability can support financial institutions in understanding risk exposure, monitoring vulnerable relationships, and making informed credit decisions\.

![Refer to caption](https://arxiv.org/html/2609.25542v1/casestudy.png)Figure 8\.Case studies, where edges \(transactions\) and corresponding neighbors \(trading partners\) influential to the prediction of the target node \(star\-shaped\) are highlighted\.

## 6\.Practical Deployment

To evaluate practical applicability,DefaultGNN’s predictions were validated in collaboration with Techfin Ratings, a licensed credit rating agency that operates a logistic regression\-based credit scoring model \(MIS\) using real\-time financial and operational data\.DefaultGNN’s predicted default scores and MIS scores were used as input features to train a separate logistic regression model, assessing whether transaction network information provides complementary value to existing credit models\. The evaluation covered 1,166,927 corporate firms and 197,397 sole proprietors using unseen data from 2018 onward, with the sequentially trainedDefaultGNNmodel from Section[5\.4](https://arxiv.org/html/2609.25542#S5.SS4)\(trained on 2018–2020\) applied without retraining\. In practice, default prediction scores are used by financial institutions to inform credit approval decisions, where firms below a given risk threshold are deemed eligible for lending\. To simulate this scenario, a risk threshold was applied to each model’s predicted scores to determine firm eligibility for credit approval\. IntegratingDefaultGNNincreased the approval rate by 6\.95 percentage points for corporate firms \(51\.57% to 58\.52%\) and 11\.11 percentage points for sole proprietors \(36\.13% to 47\.24%\), while maintaining or reducing the default rate among approved firms \(unchanged at 0\.41% for corporate firms; 0\.14% to 0\.11% for sole proprietors\)\. These results suggest that transaction network\-based risk signals complement existing credit models by expanding the pool of approvable firms without increasing portfolio risk, particularly benefiting SMEs and sole proprietors for whom traditional financial indicators are limited\.

## 7\.Conclusion

In this work, we proposeDefaultGNN, a GNN\-based framework for corporate default prediction that leverages buyer–seller transaction networks under realistic data constraints where financial statements are limited or unavailable\. Through extensive empirical analysis of large\-scale transactional data, we show that inter\-firm transactions encode critical risk signals related to transaction role and scale\. Guided by these findings,DefaultGNNmodels transactions from dual buyer and seller perspectives while incorporating transaction magnitude, consistently outperforming existing baselines — particularly for firms without prior default history\. Beyond predictive performance, the model provides interpretable insights that help identify influential trading partners, offering practical value for real\-world credit risk assessment and monitoring\. Further,DefaultGNN’s predictions complement an existing credit scoring model, greatly improving approval rates without increasing portfolio risk\.

Acknowledgements\.This work was supported by the National Research Foundation of Korea \(NRF\) grants funded by the Korea government \(MSIT\) \(RS\-2024\-00335098\) and the Ministry of Science and ICT \(RS\-2022\-NR068758\), and by Douzone/Techfin Ratings\.

## GenAI Disclosure

We acknowledge the use of LLMs \(e\.g\., GPT\-5, Claude\) for limited assistance with \(1\) editing this paper for grammar, clarity, expression variation, and length reduction to meet page limits, and \(2\) minor refactoring/debugging of code used for plotting and visualization\. All AI\-assisted edits and code changes were reviewed and validated by the authors, and all core ideas, methods, experiments, and interpretations are original contributions of the authors\.

## References

- Altman \(1968\)E\. I\. AltmanFinancial ratios, discriminant analysis and the prediction of corporate bankruptcy\.The journal of finance23\(4\),pp\. 589–609\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p1.1)\.
- Baesenset al\.\(2003\)B\. Baesens, T\. Van Gestel, S\. Viaene, M\. Stepanova, J\. Suykens, and J\. VanthienenBenchmarking state\-of\-the\-art classification algorithms for credit scoring\.Journal of the operational research society54\(6\),pp\. 627–635\.Cited by:[§5\.4](https://arxiv.org/html/2609.25542#S5.SS4.p1.1)\.
- Basel Committee on Banking Supervision \(2005\)Basel Committee on Banking SupervisionStudies on the validation of internal rating systems\.Basel Committee Working PaperTechnical Report14,Bank for International Settlements\.External Links:[Link](https://www.bis.org/publ/bcbs_wp14.pdf)Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p4.1)\.
- Beaver \(1966\)W\. H\. BeaverFinancial ratios as predictors of failure\.Journal of accounting research,pp\. 71–111\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1)\.
- Berlocoet al\.\(2021\)C\. Berloco, G\. De Francisci Morales, D\. Frassineti, G\. Greco, H\. Kumarasinghe, M\. Lamieri, E\. Massaro, A\. Miola, and S\. YangPredicting corporate credit risk: network contagion via trade credit\.PLoS One16\(4\),pp\. e0250115\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p3.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p1.1)\.
- Biet al\.\(2022\)W\. Bi, B\. Xu, X\. Sun, Z\. Wang, H\. Shen, and X\. ChengCompany\-as\-tribe: company financial risk assessment on tribe\-style graph with hierarchical graph neural networks\.InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pp\. 2712–2720\.Cited by:[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1)\.
- Boissay and Gropp \(2007\)F\. Boissay and R\. E\. GroppTrade credit defaults and liquidity provision by firms\.Technical reportECB working paper\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p3.1)\.
- Chen \(2016\)T\. ChenXGBoost: a scalable tree boosting system\.Cornell University\.Cited by:[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p2.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.6.1)\.
- Chen and He \(2012\)T\. Chen and J\. HeA network model of credit risk contagion\.Discrete Dynamics in Nature and Society2012\(1\),pp\. 513982\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p3.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p1.1)\.
- Chenet al\.\(2021\)W\. Chen, Z\. Li, and Z\. XiaoOn credit risk contagion of supply chain finance under covid\-19\.Journal of Mathematics2021\(1\),pp\. 1281825\.Cited by:[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p1.1)\.
- Chenget al\.\(2020\)D\. Cheng, X\. Wang, Y\. Zhang, and L\. ZhangRisk guarantee prediction in networked\-loans\.InIJCAI International Joint Conference on Artificial Intelligence,Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p4.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1),[§3\.2\.2](https://arxiv.org/html/2609.25542#S3.SS2.SSS2.p1.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.11.1)\.
- Crooket al\.\(2019\)J\. Crook, V\. Djeundje, R\. Calabrese, and M\. HmidCredit scoring with alternative data\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1)\.
- Engelmannet al\.\(2003\)B\. Engelmann, E\. Hayden, and D\. TascheTesting rating accuracy\.Risk16\(1\),pp\. 82–86\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p4.1),[§5\.4](https://arxiv.org/html/2609.25542#S5.SS4.p1.1)\.
- Fusteret al\.\(2022\)A\. Fuster, P\. Goldsmith\-Pinkham, T\. Ramadorai, and A\. WaltherPredictably unequal? the effects of machine learning on credit markets\.The Journal of Finance77\(1\),pp\. 5–47\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1)\.
- Ghoshet al\.\(2000\)P\. Ghosh, D\. Mookherjee, D\. Ray,et al\.Credit rationing in developing countries: an overview of the theory\.Readings in the theory of economic development7,pp\. 383–401\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p1.1)\.
- Giesecke and Weber \(2004\)K\. Giesecke and S\. WeberCyclical correlations, credit contagion, and portfolio losses\.Journal of Banking & Finance28\(12\),pp\. 3009–3036\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p3.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p1.1)\.
- Hand and Henley \(1997\)D\. J\. Hand and W\. E\. HenleyStatistical classification methods in consumer credit scoring: a review\.Journal of the royal statistical society: series a \(statistics in society\)160\(3\),pp\. 523–541\.Cited by:[§5\.4](https://arxiv.org/html/2609.25542#S5.SS4.p1.1)\.
- Hosmer Jret al\.\(2013\)D\. W\. Hosmer Jr, S\. Lemeshow, and R\. X\. SturdivantApplied logistic regression\.John Wiley & Sons\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p2.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.3.2)\.
- Huanget al\.\(2025\)Y\. Huang, C\. Nie, H\. He, Y\. Mo, Y\. Zhu, G\. Wen, and X\. ZhuMultiplex graph representation learning with homophily and consistency\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 11835–11842\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.16.1)\.
- Jacobson and Von Schedvin \(2015\)T\. Jacobson and E\. Von SchedvinTrade credit and the propagation of corporate failure: an empirical analysis\.Econometrica83\(4\),pp\. 1315–1371\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p3.1)\.
- Jaffee and Stiglitz \(1990\)D\. Jaffee and J\. StiglitzCredit rationing\.Handbook of monetary economics2,pp\. 837–888\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p1.1)\.
- Kingma \(2014\)D\. P\. KingmaAdam: a method for stochastic optimization\.arXiv preprint arXiv:1412\.6980\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p3.1)\.
- Kipf and Welling \(2016\)T\. N\. Kipf and M\. WellingSemi\-supervised classification with graph convolutional networks\.arXiv preprint arXiv:1609\.02907\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p4.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p3.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.9.2)\.
- Kumaret al\.\(2021\)J\. Kumar, V\. Kumar, D\. Verma, and S\. SharmaPrediction of corporate bankruptcy based on financial ratios using binary logistic regression\.International Journal of Statistics and Reliability Engineering7\(3\),pp\. 376–381\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1)\.
- Liuet al\.\(2025\)Y\. Liu, S\. Liu, and Y\. LuSupply chain financial risk assessment: a modified graph attention neural network\.Finance Research Letters,pp\. 108285\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p4.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p1.1)\.
- Moet al\.\(2023\)Y\. Mo, Y\. Lei, J\. Shen, X\. Shi, H\. T\. Shen, and X\. ZhuDisentangled multiplex graph representation learning\.InInternational conference on machine learning,pp\. 24983–25005\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.15.1)\.
- Ohlson \(1980\)J\. A\. OhlsonFinancial ratios and the probabilistic prediction of bankruptcy\.Journal of accounting research,pp\. 109–131\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p1.1)\.
- Parket al\.\(2020\)C\. Park, D\. Kim, J\. Han, and H\. YuUnsupervised attributed multiplex network embedding\.InProceedings of the AAAI conference on artificial intelligence,Vol\.34,pp\. 5371–5378\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.14.2)\.
- Popeet al\.\(2019\)P\. E\. Pope, S\. Kolouri, M\. Rostami, C\. E\. Martin, and H\. HoffmannExplainability methods for graph convolutional neural networks\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 10772–10781\.Cited by:[§5\.5](https://arxiv.org/html/2609.25542#S5.SS5.p1.1)\.
- Schlichtkrullet al\.\(2018\)M\. Schlichtkrull, T\. N\. Kipf, P\. Bloem, R\. Van Den Berg, I\. Titov, and M\. WellingModeling relational data with graph convolutional networks\.InEuropean semantic web conference,pp\. 593–607\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.12.2)\.
- Shumway \(2001\)T\. ShumwayForecasting bankruptcy more accurately: a simple hazard model\.The journal of business74\(1\),pp\. 101–124\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1),[§5\.4](https://arxiv.org/html/2609.25542#S5.SS4.p1.1)\.
- Stiglitz and Weiss \(1981\)J\. E\. Stiglitz and A\. WeissCredit rationing in markets with imperfect information\.The American economic review71\(3\),pp\. 393–410\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p1.1)\.
- Sun \(2024\)H\. SunResearch on financial risk assessment algorithm based on graph neural network\.InProceedings of the 2024 4th International Conference on Big Data, Artificial Intelligence and Risk Management,pp\. 932–937\.Cited by:[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1)\.
- Sundararajanet al\.\(2017\)M\. Sundararajan, A\. Taly, and Q\. YanAxiomatic attribution for deep networks\.InInternational conference on machine learning,pp\. 3319–3328\.Cited by:[§5\.5](https://arxiv.org/html/2609.25542#S5.SS5.p1.1)\.
- Thomaset al\.\(2017\)L\. Thomas, J\. Crook, and D\. EdelmanCredit scoring and its applications\.SIAM\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p1.1),[§2\.1](https://arxiv.org/html/2609.25542#S2.SS1.p2.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p4.1)\.
- Vashishthet al\.\(2019\)S\. Vashishth, S\. Sanyal, V\. Nitin, and P\. TalukdarComposition\-based multi\-relational graph convolutional networks\.arXiv preprint arXiv:1911\.03082\.Cited by:[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.13.1)\.
- Veličkovićet al\.\(2017\)P\. Veličković, G\. Cucurull, A\. Casanova, A\. Romero, P\. Lio, and Y\. BengioGraph attention networks\.arXiv preprint arXiv:1710\.10903\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p4.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1),[§5\.1](https://arxiv.org/html/2609.25542#S5.SS1.p2.1),[Table 2](https://arxiv.org/html/2609.25542#S5.T2.6.1.10.1)\.
- Wanget al\.\(2021\)D\. Wang, Z\. Zhang, J\. Zhou, P\. Cui, J\. Fang, Q\. Jia, Y\. Fang, and Y\. QiTemporal\-aware graph neural network for credit risk prediction\.InProceedings of the 2021 SIAM International Conference on Data Mining \(SDM\),pp\. 702–710\.Cited by:[§1](https://arxiv.org/html/2609.25542#S1.p4.1),[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1)\.
- Xuet al\.\(2022\)K\. Xu, Y\. Wu, H\. Xia, N\. Sang, and B\. WangGraph neural networks in financial markets: modeling volatility and assessing value\-at\-risk\.Journal of Computer Technology and Software1\(2\)\.Cited by:[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1)\.
- Yanget al\.\(2021\)S\. Yang, Z\. Zhang, J\. Zhou, Y\. Wang, W\. Sun, X\. Zhong, Y\. Fang, Q\. Yu, and Y\. QiFinancial risk analysis for smes with graph\-based supply chain mining\.InProceedings of the twenty\-ninth international conference on international joint conferences on artificial intelligence,pp\. 4661–4667\.Cited by:[§2\.2](https://arxiv.org/html/2609.25542#S2.SS2.p2.1)\.
- Yinget al\.\(2019\)Z\. Ying, D\. Bourgeois, J\. You, M\. Zitnik, and J\. LeskovecGnnexplainer: generating explanations for graph neural networks\.Advances in neural information processing systems32\.Cited by:[§5\.5](https://arxiv.org/html/2609.25542#S5.SS5.p1.1)\.