Small firms that own little physical property, such as software shops, service providers and small agricultural innovators, are routinely refused bank credit. The refusal is not driven by evidence of risk; the firm has nothing a bank knows how to seize. The most promising technical fix is to stop looking at the firm in isolation and instead read its trading relationships. A graph neural network, a model that passes information along the edges of a network so each firm's prediction draws on its customers and suppliers, can score a firm with almost no records of its own by borrowing evidence from its partners. On Ant Financial's SME transaction graph this raised default-prediction AUC from 0.574 (gradient-boosted trees) to 0.701 (Financial Risk Analysis for SMEs with Graph-based Supply Chain Mining). Its authors concede that bank examiners cannot be told why the model decided what it did. That concession matters, because lenders must send each rejected applicant an "adverse action notice" listing the specific reasons for denial. When a model borrows from neighbors, the reason may concern a third party: a firm can be refused because a downstream buyer two hops away is in distress.
We propose to make that question answerable and measurable. Every denial from a relational credit model gets decomposed exactly into contributions from the applicant's own records, from its customers, from its suppliers, and from the shape of its trading network. This tells a lender, an applicant and a regulator what fraction of a refusal was "about" the firm itself. The same decomposition also answers a policy question that current law leaves open and that different jurisdictions will settle differently: if a rule forbids denying credit on the basis of counterparty information, how many thin-file firms lose access, and how much default risk does the lender absorb? We want lenders to be able to compute this ahead of the rule change, without rebuilding the model.
Core Claim
Baseline Shapley, Shapley attribution computed against a single accepted reference applicant rather than averaged over a data distribution, was developed for independent tabular credit records (Explaining Adverse Actions in Credit Decisions Using Shapley Decomposition). It extends naturally to graph models if the "players" are defined as information sources rather than individual features. Let $f$ be the model's default log-odds for a firm computed on its local network neighborhood, let $x^D$ be the declined firm with its neighborhood, and let $x^A$ be a reference accepted firm with its neighborhood. We write
$$f(x^D) - f(x^A) = \sum_{g \in G} \phi_g ,$$
where $G$ is a small set of source groups and $\phi_g$ is the Shapley value of group $g$. The groups are the applicant's own cash-flow blocks, its downstream customers' features, its upstream suppliers' features, two-hop partners' features, and the applicant's edge set. Each $\phi_g$ is computed by swapping that group's values between $x^D$ and $x^A$ across all coalitions of the other groups. With roughly ten groups this takes about a thousand forward passes of the model on a small subgraph per applicant. The computation is therefore exact, sums to the true score gap, and avoids the intractability that On the Hardness of Computing Counterfactual and Semifactual Explanations in XAI proves for optimal counterfactual search. That hardness concerns searching input space for a minimal change. We only evaluate a fixed, small set of swaps.
From the decomposition we derive two quantities that no paper in our file measures. The first is the borrowed-denial rate: the share of denials whose largest attribution group lies outside the applicant's own records. The second is the permissibility cost of a candidate rule. For a rule such as "counterparty information may raise a price but may not cause a refusal," the permissibility cost is the change in approvals, approvals among thin-file firms, and realized default rate at a fixed approval budget when the forbidden groups are pinned to reference values at decision time.
The novelty is not a new predictor. It is an accountability layer that says whose data drove a commercial-credit denial, plus a way to cost out unsettled rules before they are written. The design responds to regulatory lag. U.S. states are enacting AI laws far faster than the federal government, but through committee bottlenecks that make timing unpredictable (The Bottleneck in AI Governance and Oversight: Evidence from 1,419 State Bills). Across six major jurisdictions only one of fifteen governance processes is aligned (Navigating Governance Paradigms). A lender serving many jurisdictions cannot wait for one answer on third-party data. It needs to know in advance what each answer would cost and whom it would exclude.
Research Program
The prior work plays three roles here. Motivating work established that relational evidence matters for exactly the firms the financing gap concerns. Yang et al. show that only 1.5% of SMEs had patent data directly but 22.7% did once one-hop partners were included, that firms with more than ten partners default at under half the average rate, and that downstream ties buffer risk more than upstream ties (Financial Risk Analysis for SMEs). Bankruptcy models that separate a firm's internal fragility from shocks spreading through its network make the same point from the distress side (Combining intra-risk and contagion risk for enterprise bankruptcy prediction using graph neural networks).
Method-supplying work gives us the attribution machinery and the constraints that keep it stable. B-Shap supplies the reference-point decomposition and the evidence that monotone, grouped models give stable reasons (Explaining Adverse Actions). Wang et al. supply the production template of monotone XGBoost plus TreeSHAP mapped to regulatory reason codes for small-business bank data (Using Small Business Banking Data for Explainable Credit Risk Scoring). Both treat each borrower as an isolated row, and we extend them to neighborhoods.
Constraining work rules out certain designs. Explanations fitted as separate surrogate models can be tuned to look fair while the underlying model is not (Fairwashing: the risk of rationalization), so we attribute the deployed model itself and never a surrogate. B-Shap's own open problem is that changing the reference applicant changes the reasons. That problem is worse on graphs, where the reference also needs a neighborhood, so reference sensitivity becomes a primary measurement rather than a footnote.
Concretely, we will build four things. The first is a relational scorer: a monotone message-passing network whose aggregation over neighbors is constrained to be monotone in signed risk features, following the regularization evidence in Wang et al. and B-Shap. The second is the grouped attribution engine described above. Within the own-records group it maps sub-attributions to standard reason codes using the Weight-of-Evidence mapping of Wang et al. The third is a reference protocol comparing three reference choices: a median accepted firm with its actual neighborhood; the applicant's own network with neighbors set to accepted-firm values, which isolates "what if your partners were healthy"; and the applicant stripped of edges. The fourth is a scenario module that encodes candidate legal rules as sets of groups that may not drive a denial and reports permissibility cost under both inference-time pinning and full retraining without those groups.
Baselines
The naive reference point is the do-nothing compliance choice: a logistic scorecard on the applicant's own application and cash-flow features, with no relational data. It is trivially explainable and represents what a cautious lender does today. On isolated bank-statement data the reported gain from moving beyond application-only features is a 24.6% relative AUROC improvement, reaching 0.806 on 611 Malaysian loans (AI-BAAM). Wang et al. report a 7% K-S improvement of monotone XGBoost over a traditional scorecard on QuickBooks Capital data. We will refit the scorecard with the same feature construction and tuning budget as every other model. It defines the zero point for permissibility cost.
The strongest own-records competitor is monotone XGBoost with TreeSHAP reason codes, as deployed by Wang et al. It is fully auditable and production-proven. Wang et al. report that monotone constraints alone raised K-S by 8% over unconstrained XGBoost and that its reasons agreed with legacy scorecard reasons 42–52% of the time. Any relational model must beat it on K-S and AUC at matched hyperparameter search budgets on identical train/test splits, split by time so that later loans are never used to predict earlier ones.
The strongest relational competitor is the unconstrained spatio-temporal graph model of Yang et al., whose reported default AUC on the Ant graph is 0.701, against 0.671 for a static graph attention network and 0.574 for gradient-boosted trees. We will reimplement it on our data rather than compare with numbers from a different graph. Our monotone relational scorer counts as competitive if it stays within roughly the gap separating the static and temporal graph models in that paper, about 0.03 AUC, of the unconstrained graph model, while producing exact source attributions.
The ablation that isolates our key ingredient is the same relational model explained with per-feature attributions over the flattened neighborhood instead of source groups. This tests whether grouping by information source, rather than Shapley on graphs per se, is what produces stable, legally usable reasons. The comparison uses the same model, forward-pass budget and reference.
Why not just…?
Why not just hand-engineer a few relational aggregates and feed them into monotone XGBoost? This is the most serious threat. Yang et al.'s own stylized facts, partner count and the downstream/upstream split, suggest a handful of features might capture much of the graph model's lift, and TreeSHAP would then explain them exactly. We do not argue around this. Our attribution and permissibility analysis applies unchanged to such a model, because each engineered aggregate belongs to a source group. The experiments include this flattened model as a first-class candidate. If it matches the graph model, the conclusion is that the accountable relational lender should use it. The borrowed-denial rate and permissibility cost remain the new measurements either way.
Why not just drop all counterparty data and avoid the question? That is the do-nothing baseline. Its cost is what we measure, and the receptive-field evidence (patent coverage rising from 1.5% to 22.7% through partners) suggests the cost falls on precisely the thin-file, asset-light firms the financing gap concerns. The scenario experiments report approvals among firms with sparse own records under this rule, so the trade-off becomes a number rather than an assumption.
Why not just issue counterfactual explanations ("what would need to change for approval")? Optimal counterfactuals are NP-hard and inapproximable for tree ensembles and ReLU networks (On the Hardness of Computing Counterfactual and Semifactual Explanations). Global rule-based recourse such as T-CREx assumes freely adjustable static features (Counterfactual Metarules for Local and Global Recourse), and a firm cannot change its customers' health. Recourse also answers a different question from the one adverse action law asks, which is what the reasons were. We see source attribution as a prerequisite for sensible recourse: only own-record groups are candidates for action.
Why not just fit an interpretable surrogate or rule summary of the graph model? Rule summaries with guaranteed global consistency exist for tabular credit data (Globally-Consistent Rule-Based Summary-Explanations; An Interpretable Model with Globally Consistent Explanations for Credit Risk). Their integer programs grow expensive in high dimensions, though, and any surrogate opens the door to the rationalization LaundryML demonstrates, which reached 90.8% fidelity with sharply reduced apparent unfairness. Our decomposition sums exactly to the deployed model's score gap. That property can be checked by an auditor, which a surrogate's fidelity cannot guarantee.
Proposed Experiments and Decisive Checks
No public dataset on Lacuna combines SME loan outcomes with a transaction graph. The Ant graph and QuickBooks data are proprietary, and the AI-BAAM data is a single-firm 611-loan sample. The program therefore runs at two levels.
The first level is a semi-synthetic benchmark where the true cause of each default is planted and known. Firm-level cash flows follow a simple autoregressive process. Defaults arise from a mix of own fragility and shocks spreading through the network, in the intra-risk/contagion split of the bankruptcy-GNN work. The generator is calibrated to match the reported stylized facts: sparse own-feature coverage near 1.5%, a default rate halved above ten partners, and downstream ties buffering more than upstream ones. Its decisive check is attribution recovery: the fraction of denials whose top attribution group matches the planted cause. Source-grouped attribution should beat flattened per-feature attribution and TreeSHAP on engineered aggregates. If it does not, the grouping idea fails.
The second level is a lender partnership with a real counterparty graph built from bank-statement fund flows of the kind AI-BAAM already extracts. On these data we report AUC and K-S against the baselines above. We also report the borrowed-denial rate, and reference sensitivity measured as rank agreement (Kendall's tau) of top reason groups across the three reference choices. Finally, for each rule scenario, we report permissibility cost at a fixed approval rate, broken out for thin-file firms. The headline claim depends on borrowed denials being common, and that is uncertain. If they turn out to be rare, own-record explanations suffice and the legal worry is moot for this data.
The second decisive check is whether inference-time pinning of forbidden groups tracks full retraining. If pinning over- or understates the permissibility cost relative to a model retrained without those groups by more than the variation across random seeds, the claim that rules can be costed without rebuilding fails, and the scenario module must retrain.
The third decisive check is reference stability. If top reason groups reorder under reasonable reference choices for a large share of applicants, a source-attributed notice is not yet defensible, and the reference must be fixed by convention rather than derived. Determining which of these outcomes holds is itself the result a regulator needs before writing the rule.
References
- Financial Risk Analysis for SMEs with Graph-based Supply Chain Mining
- Explaining Adverse Actions in Credit Decisions Using Shapley Decomposition
- On the Hardness of Computing Counterfactual and Semifactual Explanations in XAI
- The Bottleneck in AI Governance and Oversight: Evidence from 1,419 State Bills
- Navigating Governance Paradigms: A Cross-Regional Comparative Study of Generative AI Governance Processes & Principles
- Combining intra-risk and contagion risk for enterprise bankruptcy prediction using graph neural networks
- Using Small Business Banking Data for Explainable Credit Risk Scoring
- Fairwashing: the risk of rationalization
- AI-BAAM: AI-Driven Bank Statement Analytics as Alternative Data for Malaysian MSME Credit Scoring
- Counterfactual Metarules for Local and Global Recourse
- Globally-Consistent Rule-Based Summary-Explanations for Machine Learning Models: Application to Credit-Risk Evaluation
- An Interpretable Model with Globally Consistent Explanations for Credit Risk