DefaultGNN: A Dual-Perspective GNN Framework for Predicting Corporate Default from Buyer-Seller Transaction Networks

• For SMEs and sole proprietors, missing, outdated, or incomplete financial statements can create information asymmetry, and traditional scoring tends to overestimate their risk.Source: §1 · p. 1
• Existing GNN approaches often simplify transactions into unidirectional, binary links, overlooking bidirectional exposure and differences in transaction scale.Source: §1 · p. 2
• DefaultGNN combines multiplex buyer-view and seller-view transaction networks to model corporate default risk, particularly for firms with limited intrinsic financial signals.Source: §1 · p. 2
• The study analyzes six years of electronic tax-invoice data and reports that DefaultGNN complements an existing credit score, improving approval rates without increasing default risk.It also presents network visualizations of how distressed trading partners contribute to predicted risk.Source: §1 · p. 2

2.1 Modeling for Corporate Default Prediction

• Financial-statement models and attribute-based machine-learning methods depend on firm-level data and generally do not model inter-firm relationships.Source: §2.1 · p. 2

2.2 Network-Based Financial Risk and GNNs

• Although GNNs have been applied to financial networks, many existing approaches use non-transactional relations or single-view, unweighted, unidirectional transaction graphs.The paper combines transaction-data analysis with role- and scale-aware network modeling.Source: §2.2 · p. 2

• The dataset spans 2018–2023 and combines firm attributes and transactions from Douzone Bizon’s electronic tax-invoice system with default labels from a licensed credit rating agency.Source: §3.1 · p. 2
• Preprocessing removes firms with invalid identifiers or missing business type and discards non-positive transaction amounts.Source: §3.1 · p. 2
• Transactions involving partners with prior default history are described as more likely to precede future default of the target firm.Source: Figure 1 · p. 3
• Firms with defaulted trading partners are described as having higher future default rates than firms whose partners have no prior default history.Source: Figure 2 · p. 3
• The Figure 3 caption describes later-defaulting firms as having more defaulted trading partners and notes monthly, role-based aggregation of directed interactions.The caption states that at most two directed interactions per firm pair are retained, with transaction amounts summed.Source: Figure 3 · p. 3

3.2 Default-Transaction Correlation

• Transactions involving partners with a default in the previous 12 months are consistently associated with a higher likelihood of target-firm default in the next 12 months.Source: §3.2.1 · p. 3

3.2.1 Effect of Partner Default.

• Firms that transact with at least one defaulted partner have higher future default rates, and firms that later default tend to have more and a greater proportion of defaulted partners.These patterns remain consistent across time.Source: §3.2.1 · p. 3

3.2.2 Effect of Transaction Role.

• Sales to and purchases from defaulted partners are both associated with elevated future default risk, supporting distinct buyer and seller transaction views.Source: §3.2.2 · p. 3, Figure 4 · p. 3

3.2.3 Effect of Transaction Scale.

• Large-scale transactions with defaulted partners are associated with significantly greater future default likelihood than no such transactions, while small-scale transactions have a much weaker effect.Transaction scale is the transaction amount divided by the firm’s total transaction volume within the same month; the figure caption defines large scale as proportion ≥0.5 and small scale as proportion <0.5.Source: §3.2.3 · p. 3, Figure 5 · p. 3

3.2.4 Summary of Analysis.

• Partner default history, buyer-seller role, and transaction scale are identified as factors in bidirectional, heterogeneous default propagation, with economically significant transactions contributing disproportionately.Observed trends remain consistent across the years considered.Source: §3.2.4 · p. 3
• The findings motivate dual-perspective, edge-weighted graph modeling that accounts for heterogeneous risk exposure.Source: §3.2.4 · p. 3

3.3 Multiplex Transaction Networks

• The two network views encode different exposure directions: buyer stress can propagate to sellers, while seller delivery failures or disruptions expose buyers.The seller-view edge points from buyer to seller; the buyer-view edge reverses that direction.Source: §3.3 · p. 4
• Each edge weight is the transaction amount relative to the corresponding firm’s monthly sales in the seller view or purchases in the buyer view.Weights may be scaled to mitigate the effect of extreme values.Source: §3.3 · p. 4
• The representation comprises two directed, edge-weighted graphs with a shared node set and distinct transactional semantics, jointly capturing role- and scale-dependent risk patterns.For each year, firms are nodes and transactions are directed edges.Source: §3.3 · p. 4

3.4 Firm Attributes

• Each firm is represented by 108 non-financial attributes for settings where detailed financial statements are unavailable.The attributes capture basic operational characteristics under realistic data constraints.Source: §3.4 · p. 4
• The attributes include a recent default indicator, taxation category, business-type encodings at multiple granularities, and log-scaled aggregate sales and purchase statistics.The aggregate statistics are computed from historical transaction records.Source: §3.4 · p. 4
• These features omit specific trading-partner and network information, providing an attribute-only baseline enhanced by relational transaction-network information.Source: §3.4 · p. 4

4.1 Problem Setting and Overview

• DefaultGNN learns buyer- and seller-view representations, fuses them through role-adaptive gating, and uses the result to predict future defaults.View consistency regularization stabilizes this integration.Source: §4.1 · p. 4, Figure 6 · p. 4

4.2 View-Specific Graph Encoders

• Parallel GNN encoders aggregate trading-partner information using transaction-scale edge weights to capture role-specific risk propagation.The normalized aggregation coefficient depends on both graph structure and transaction scale.Source: §4.2 · p. 4

4.3 Role-Adaptive Gated Fusion

• A per-firm gate sets the relative importance of seller- and buyer-view embeddings, which are concatenated for default prediction.View-consistency regularization encourages compatible risk estimates when both views are confident while allowing divergence when either is uncertain.Source: §4.3 · p. 5

4.4 Training Objective

• Training combines weighted binary cross-entropy with a view-consistency penalty, with the classification weight emphasizing correct identification of defaulted firms.The regularization strength is controlled by λ.Source: §4.4 · p. 5

5.1 Experimental Setup

• Experiments use six yearly datasets, stratified 70/10/20 train-validation-test splits, four baseline categories, and Accuracy Ratio derived from AUC.DefaultGNN is compared with attribute-based, standard GNN, multi-relational GNN, and multiplex models; AR is defined as 2 × AUC − 1.Source: §5.1 · p. 5

5.2 Performance Comparison (RQ1)

• DefaultGNN significantly outperforms the best baseline on every dataset, while attribute-based methods trail graph models and DefaultGNN is especially advantageous for firms without prior default history.Multiplex models outperform standard and multi-relational GNNs; the authors associate DefaultGNN’s additional gains with role-adaptive gated fusion.Source: Table 2 · p. 6, §5.2 · p. 6

5.3 Ablation Study (RQ2)

• The full DefaultGNN performs best: collapsing buyer and seller views causes larger drops than using either view alone, and removing transaction weights drops performance below single-view models.Concatenation and removing the consistency regularizer also reduce performance; cross-attention causes a still larger decline. NoHist AR remains 0.452–0.472 across tested hyperparameter choices on the 2021 dataset.Source: §5.3 · p. 6, Table 3 · p. 6

5.4 Out-of-time Analysis (RQ3)

• DefaultGNN outperforms GCN and XGBoost in both out-of-time configurations, maintaining performance as the training-test gap reaches five years and benefiting from sequential retraining.The configurations train on 2018 alone or sequentially on 2018–2020, then test on later years.Source: §5.4 · p. 7, Figure 7 · p. 7

5.5 Case Studies (RQ4)

• Two case studies show DefaultGNN identifying indirect risk through a previously defaulted partner and capturing a signal spanning buyer- and seller-view transactions that single-view GCN misses.The visualizations attribute predictions to influential transaction edges and trading partners.Source: §5.5 · p. 7, Figure 8 · p. 7

6 Practical Deployment

• The evaluation combines DefaultGNN and MIS scores as inputs to a separate logistic regression model, using unseen data from 2018 onward without retraining DefaultGNN.It covers 1,166,927 corporate firms and 197,397 sole proprietors; the existing MIS model uses real-time financial and operational data.Source: §6 · p. 7
• Approval rates rose 6.95 percentage points for corporate firms and 11.11 percentage points for sole proprietors when DefaultGNN was integrated.Corporate approvals increased from 51.57% to 58.52%; sole-proprietor approvals increased from 36.13% to 47.24%.Source: §6 · p. 7
• Among approved firms, default rates remained 0.41% for corporate firms and fell from 0.14% to 0.11% for sole proprietors.The authors conclude that transaction-network signals complement existing credit models without increasing portfolio risk.Source: §6 · p. 7

7 Conclusion

• DefaultGNN models corporate transactions from buyer and seller perspectives while incorporating transaction magnitude, and outperforms existing baselines, particularly for firms without prior default history.The framework is motivated by settings where financial statements are limited or unavailable.Source: §7 · p. 7
• The model provides interpretable insights that identify influential trading partners for credit-risk assessment and monitoring.The conclusion also notes that DefaultGNN predictions complement an existing credit scoring model.Source: §7 · p. 7
• Integrating DefaultGNN predictions with an existing credit scoring model improves approval rates without increasing portfolio risk.Source: §7 · p. 7