FraudGNAM: Inherently Interpretable Spectral GNN for Graph Fraud Detection

Abstract

Graph neural networks for fraud detection face a fundamental tension: spectral methods that capture heterophilic patterns sacrifice interpretability, while interpretable models lack the spectral capacity to detect sophisticated fraud. We introduce FraudGNAM (Fraud Graph Neural Additive Model), the first inherently interpretable spectral GNN with learnable multi-scale filters that resolves this tension. FraudGNAM decomposes the fraud score into individually auditable components—per-feature contributions via neural additive models, pairwise feature interactions, multi-scale spectral band responses via learnable polynomial filters, and explicit feature-spectrum interactions—all combined through a transparent scalar-sum architecture. Each component’s contribution to the final prediction is independently quantifiable, unlike post-hoc explanation methods that approximate black-box decisions. On seven GADBench datasets spanning homophily ratios from 0.60 to 0.98, FraudGNAM achieves state-of-the-art AUROC on five of seven datasets against SEC-GFD, the strongest black-box baseline, while winning 14 of 21 metric-dataset comparisons overall. An adaptive configuration mechanism selects spectral order, high-pass filter count, and contrastive regularization based on measurable graph properties, eliminating per-dataset hyperparameter search.