This study examines how new money laundering patterns can be spotted in financial transaction networks without depending on past labels, using a three-stage, temporally safe machine learning setup in the USA. We started by building time-resolved transaction graphs from the Elliptic dataset, making sure that no model ever saw information from the future during training or testing. From these graphs, we created node embeddings and ran several unsupervised anomaly detection methods, each offering its perspective on what might look suspicious. This first step gave us a baseline for how each method performed over time. The next step was to bring their outputs together in a rank-based ensemble. The goal was to even out the biases and blind spots of individual detectors and make the identification of high-risk transactions more consistent. In the final stage, we clustered the ensemble’s top-ranked results, grouping related nodes into candidate typologies that reflected both structural and behavioral patterns in the network. What came out of this was a detection pipeline that outperformed any single method and did a better job of surfacing coherent, interpretable clusters of suspicious activity. One of the main takeaways is that emerging laundering methods rarely hinge on a single, obvious signal. They are built from a mix of interactions, small shifts in structure, timing, and connectivity, that only stand out when viewed from multiple angles over time. This shifts AML from relying on static detection lists toward a system that evolves with the network, giving analysts a practical way to uncover threats they have never encountered before while staying true to real investigative timelines.
Mir Mohtasam Hossain Sizan (2025) studied this question.