We introduce GRAFID, a decision framework that helps financial institutions determine whether graph-based methods justify their computational overhead for transaction fraud detection. Through twenty-plus model configurations across two public datasets (IEEE-CIS and European Credit Card), we demonstrate that the value of graph structure depends primarily on feature richness. On the feature-rich IEEE-CIS dataset, XGBoost achieves AUPRC of 0.508, outperforming GraphSAGE (0.411) at roughly one-sixtieth the cost. On the feature-sparse Credit Card dataset, the two approaches perform identically. GRAFID provides three actionable indicators: Feature Richness Index, Graph Signal Gain, and Cost-Effectiveness Ratio, enabling practitioners to make informed investment decisions before committing to graph infrastructure.
Karl J. Mollan Neyra (Fri,) studied this question.