Authors - Srikumar Nayak Abstract - Anti–money laundering (AML) monitoring is difficult because suspicious behavior is rarely a single abnormal transaction; it is usually a short sequence of linked transfers across many entities. Standard tabular models miss these links and often produce alerts that are hard to justify during review. To address this, we propose GraphAML-X, a practical pipeline that turns raw transaction logs into a knowledge graph and produces case-level evidence for analysts. The main issue we target is fragmented identity (the same actor appearing under noisy identifiers) and weak case explanations (high scores without clear paths or rule triggers). GraphAML-X first performs entity resolution to merge duplicate accounts and identifiers using rules plus a learned match score, so the graph represents real actors. It then learns temporal graph embeddings from the timeordered transaction network to capture multi-hop laundering patterns such as rapid circulation and hub–spoke behavior. Finally, it combines graph risk with rule-hybrid case reasoning: regulatory red-flag rules propose candidate alerts, and the graph model ranks them while emitting audit-ready evidence (top subgraph paths, key neighbors, and triggered rules) and alert-volume control via a calibrated threshold. Using the Micro-AmlSim dataset, GraphAML-X achieves an AUC-ROC of 0.982 and an AUC-PR of 0.741, improving the strongest baseline GNN by +0.034 AUC-PR. At a fixed alert rate of 1% of transactions, it attains 0.686 recall while reducing false alerts by 18.9% compared to rule-only screening. These results show that GraphAML-X can improve detection while producing reviewable and policy-aligned AML cases.
Open Zoom