Guilt by association: Interpretable multiplex graph learning of accounting-risk contagion

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Severe accounting failures-material internal-control weaknesses, fraudulent restatements, and enforcement actions-are rare, costly, and relational: they cluster along the networks binding firms to common auditors, directors, and owners. We recast next-year failure prediction as node classification on a temporal multiplex firm graph whose layers encode a nested auditor hierarchy, board interlocks, and common ownership, and ask when the graph adds to the node-attribute signal. Methodologically, the usual graph-versus-tabular comparison is unfair: graph models receive node features that the tabular baselines are denied; once equalised, every standard graph network falls behind gradient-boosted trees. Architecturally, we introduce ACE-GNN, a pure multiplex network whose strength comes not from a richer operator-ruled out by a deliberate negative result-but from quantile encoding of node attributes and zero-inflated neighbour-exposure signals. ACE-GNN matches the strongest tabular learner, and a tree handed the same exposures attains near-identical accuracy, so the graph's advantage lies in faithful relation-level explanation, not accuracy. For interpretability, silent failures-firms whose warnings arrive only through their ties-are traced by edge-removal counterfactuals to specific auditor peers. These results delineate when graph-based AI earns its place on relational-but-tabular problems-and when it does not.

키워드

Graph-based AI; Multiplex networks; Explainable AI; Graph neural networks; Numerical feature embeddings; Accounting-risk contagion; Early warning; FINANCIAL RATIOS; PREDICTION
제목
Guilt by association: Interpretable multiplex graph learning of accounting-risk contagion
저자
Kang, Hojun; Lee, Sang-Gun
DOI
10.1016/j.ins.2026.124208
발행일
2027-02
유형
Article
저널명
Information Sciences
권
762