CFGN: causal flow graph networks for causally-aware fraud detection in imbalanced transaction graphs

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Fraudulent activities in financial, e-commerce and digital transaction networks are getting difficult to detect, as fraudsters use the network’s legitimate architecture to hide their actions, and fraud activity is a small minority of all transactions. Existing graph neural network (GNN) detectors are unreliable due to this mix of extreme class imbalance, adversarial camouflage, and structural heterogeneity across application domains, because they only accumulate neighbour information via correlation and are unable to tell the difference between a genuine, causal link and a link that has been crafted for benign purposes but nonetheless camouflages the connection. We propose the Causal Flow Graph Network (CFGN), a graph neural network framework that learns both local and global structural effects, and explicitly removes the non-causal confounded neighbour signals before they are propagated. At its heart stands a causal gate that formally approximates Pearl’s backdoor adjustment criterion from structural causal modelling, which suppress the presence of confounded, camouflage-driven neighbouring causal signals during message passing and only consider causally identified signals to inform the representation of the target node. This causal gate is combined with structural attention through a learned adaptive mixture coefficient \(\lambda ,\) which automatically calibrates, per dataset, how much the model should rely on causal filtering versus standard attention. We apply CFGN to three application domains that differ in fraud type, graph topology, and labeling authority: cryptocurrency transactions (the Elliptic Bitcoin network), anti-spam review detection (YelpChi), and e-commerce behavioural fraud (Amazon reviews). Across all three, the learned mixture coefficient \(\lambda \) exceeded the 0.5 dominance threshold on the two domains with verified, adversarially rich labels (Bitcoin: \(0.71 \pm 0.02;\) YelpChi: \(0.73 \pm 0.04)\) and fell below it on the heuristically labeled Amazon domain \((0.39 \pm 0.03)\) , an empirical pattern consistent with causal gating becoming dominant specifically where adversarial camouflage is present. CFGN achieves a cross-validated AUC-PR of 0.6567 and AUC-ROC of 0.8868 on the professionally labeled Elliptic Bitcoin dataset, outperforming seven baseline and state-of-the-art GNN models on precision-recall metrics. In addition to raw accuracy, CFGN provides significantly higher robustness under the adversarial camouflage edge injection setting (with \(0.4 \pm 0.0\) injection rate it achieves an average precision of \(75.0\%\) on YelpChi while GAT achieves \(50.0\%\) and GCN achieves \(53.0\%)\) ; and under the temporal concept drift setting, CFGN’s performance drops by \(8.6\%\) while GCN’s drops by \(53.6\%.\) The combination of these results demonstrates that the explicit modeling of causal structure in message passing leads to improvements in both detection accuracy and material increases in adversarial or non-stationary robustness to real-world fraud detection.