DefaultGNN: A Dual-Perspective GNN Framework for Predicting Corporate Default from Buyer-Seller Transaction Networks
Junghoon Kim, Hyunsung Kim, Seungyoon Choi, KyoungYong Park, Jihun Lee, YongGu Ji, Chanyoung Park
cs.LG · cs.AI
Quick verdict
Claims dual buyer/seller-view GNN lifts corporate default prediction; large Korean tax-invoice study is solid but proprietary-data-bound with modest gains.
Summary
This CIKM '26 paper tackles corporate default prediction when financial statements are sparse, as is common for SMEs. Using six years (2018-2023) of Korean electronic tax-invoice records from one vendor, covering hundreds of thousands to about a million firms and tens of millions of directed transactions per year, the authors first show empirically that having a defaulted trading partner raises a firm's future default risk in both directions (sales to and purchases from defaulted partners), and that large-scale transactions contribute disproportionately. They build multiplex buyer-view and seller-view weighted directed networks and propose DefaultGNN: two parallel GCN encoders produce view-specific embeddings, fused through a role-adaptive gate (gate values correlate r=0.73 with firms' sales/(sales+purchases) balance) plus a confidence-weighted view-consistency regularizer, trained with weighted BCE to predict next-year default. Across six yearly datasets, measured by Accuracy Ratio (2*AUC-1), DefaultGNN beats attribute-based, standard GNN, multi-relational GNN, and multiplex baseline families, most strongly on firms without prior default history; ablations show edge weights, gating, and the regularizer each matter, and out-of-time tests are encouraging. In deployment with a licensed rating agency, approval rates rose roughly 7 (corporates) and 11 (sole proprietors) percentage points without raising default rates. Main limitations: proprietary single-vendor data, random within-year splits that permit transductive leakage, and interpretability supported only by two qualitative case studies.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Default history of trading partners, buyer/seller role, and transaction scale are three empirically separable drivers of future corporate default (Section 3, Figs. 1-5).
- Explicitly modeling buyer-view and seller-view networks with scale-weighted edges and role-adaptive gated fusion outperforms attribute-based, standard GNN, multi-relational, and multiplex baselines on all six yearly datasets (Table 2, Wilcoxon p=0.016).
- The advantage is largest for firms without prior default history (NoHist), the population where financial-statement models fail (Section 5.2).
- Integrated with a licensed agency's existing scoring model, DefaultGNN raises approval rates by 6.95pp for corporates and 11.11pp for sole proprietors without increasing default rates among approved firms (Section 6).
Strengths
Empirical analysis genuinely motivates the design
Section 3 (Figs. 1-5) shows partner default raises future default bidirectionally and that large-scale (proportion ≥ 0.5) transactions with defaulted partners matter far more than small ones; the dual-view, edge-weighted architecture follows directly from these findings rather than being post-hoc.
Well-targeted ablations on a large dataset
Table 3 (2021) isolates single views (0.544/0.545 vs 0.583 full), collapsed views (0.531/0.516), edge-weight removal (0.535 — below even single-view), gating (0.565), the consistency regularizer (0.560), and cross-attention (0.537), so each design claim is backed by a specific variant.
Industry-standard metric and a practically motivated subset
Deployment validation beyond offline AUC
Honest separation of attribute features from network signal
Weaknesses
Fully proprietary single-vendor data
Evaluation design favors transductive leakage
Deployment comparison lacks a controlled baseline
Interpretability is only anecdotal
Baseline fairness is uncertain
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DefaultGNN (CIKM '26, full paper) predicts next-year corporate default from buyer-seller transaction networks, motivated by the unreliability of financial statements for SMEs. The paper makes three empirical claims (Section 3): partner default history, transaction role (buyer vs seller), and transaction scale all shape future default; and two system claims: a dual-view, scale-weighted GNN outperforms four baseline families (Table 2), and its scores add measurable value to a production credit model (Section 6). The dataset is unusually strong — six years of Korean electronic tax invoices with Basel II default labels — covering 0.42M-1.0M firms and 14M-55M directed transactions per year (Table 1), with default rates of 0.79%-1.47%.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DefaultGNN?
DefaultGNN is a graph neural network framework for predicting corporate default from buyer-seller transaction networks. It processes two directed, edge-weighted views of the same firms (a buyer view and a seller view built from electronic tax-invoice data), learns embeddings with two GCN encoders, and fuses them with a role-adaptive gate plus a view-consistency regularizer before predicting whether a firm will default in the following year.
What data was DefaultGNN trained on, and is it public?
Six years of Korean electronic tax-invoice records (2018-2023) from the vendor Douzone Bizon, covering roughly 0.42-1.0 million firms and about 14-55 million transactions per year, with default labels defined by Basel II criteria from the licensed rating agency Techfin Ratings. The dataset is proprietary; only source code is released on GitHub.
How much better is DefaultGNN than simpler models?
Using the Accuracy Ratio metric (2·AUC−1), DefaultGNN scores 0.542-0.605 across the six years versus about 0.50-0.59 for the strongest multiplex-graph baselines (e.g., 0.583 vs 0.550 for DMGI on 2021), with the largest gains on firms without prior default history (NoHist). Gains over the best baselines are roughly 0.02-0.04 AR, consistent across years.
Did the model work in a real credit-setting deployment?
The authors report a deployment study with the licensed rating agency Techfin Ratings on about 1.17 million corporate firms and 197 thousand sole proprietors: adding DefaultGNN's score to the agency's logistic-regression scoring model raised approval rates by about 7 percentage points for corporates (51.57% to 58.52%) and 11 for sole proprietors (36.13% to 47.24%), while default rates among approved firms stayed flat or fell slightly (0.41% unchanged; 0.14% to 0.11%).
Can I reproduce the paper's results?
Exact replication is currently limited: the transaction and default data are proprietary to a Korean vendor-agency pair, so independent verification of Table 2 would require data access. The code repository may support architectural study, and the method's core ideas could be reimplemented on public transaction-network datasets, but reported numbers themselves are unlikely to be reproduced externally.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure values (Figs. 1-5, 7, 8) are not recoverable from the text extraction; only captions and axis ranges are visible, so OOT degradation magnitudes could not be audited.
- The text appears complete for Sections 1-7 with tables intact, so confidence is high despite the figure limitation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.