Jump-Attentive Graph Learning for Camouflaged Financial Fraud in Transaction Networks

This paper presents a finance-centric graph learning approach that embeds fraud detection directly into transaction networks through a two-pronged design: a similarity-aware neighbourhood sampler and a jump-attentive GNN. The sampler leverages relationship-specific covariance structure and mutualneighbor weighting to surface look-alike anchors while injecting controlled randomness to preserve rare but decisive financial risk signals. The proposed network combines multi-relational attention over sampled neighborhoods with depth-bounded jump connections that retain distant indicators such as chargebacks without succumbing to over-smoothing. Evaluated on a largescale, highly imbalanced e-commerce transaction graph and an external review-fraud benchmark, the method improves AUC and recall over strong baselines, demonstrating robust identification of camouflaged financial fraud in production-style settings. The results highlight a practical CS-for-finance pathway that unifies graph representation learning, imbalanced risk modeling, and deployable data engineering for modern payment ecosystems.