Many high-impact abuses are relational rather than isolated events: Accounts share devices, sellers share payout instruments, and coordinated groups reuse infrastructure. This chapter presents graph-based modeling as a framework for capturing those connections across buyers, sellers, devices, addresses, and financial entities. It describes entity linkage and graph construction and then develops two complementary approaches: engineered graph features (neighborhood aggregates and structural patterns) and learned representations via graph neural networks. Graph methods are shown to be especially effective for collusion rings, synthetic identity networks, mule activity, and repeated linkages that evade single-event scoring. Operational feasibility is emphasized, including scalability, entity resolution accuracy, temporal graph updates, and adversarial adaptation. The chapter also connects graph modeling to enforcement workflows, where the unit of action is often the network rather than the individual account. Practical integration topics include temporal edges, cold-start entities, and how graph signals can be surfaced as features or scores within existing decision engines and analyst tooling. The chapter emphasizes that linkage quality often dominates model performance, and provides guidance on building trustworthy entity graphs. By incorporating relational context, graph-based models turn isolated alerts into coherent investigative narratives and enable earlier intervention against coordinated exploitation.