What is it about?
Detecting illicit actors on financial networks, such as those committing money laundering or fraud, is hard because true examples are rare and expensive to verify. Machine learning can help, but training a good detector needs many labeled examples. Active learning is a way around this: instead of labeling many examples blindly, the model asks an expert to label only the most informative samples, reducing labeling costs while producing a highly effective model. We built C2AL, an active-learning method that picks which actor an expert should review next so the model learns as quickly as possible. C2AL trains two classifiers side by side: one that looks at each actor in isolation, and one that also looks at the actor's neighbors in the transaction graph. When the two classifiers disagree on an actor's label, the network around that actor has changed the answer, and that actor is exactly the kind of high-value example to send to the expert for review. We tested C2AL on two real Bitcoin networks (Elliptic++) and four synthetic anti-money-laundering benchmarks (AMLworld).
Featured Image
Photo by Mehdi Mirzaie on Unsplash
Why is it important?
Catching illicit actors on financial networks matters for fraud prevention, sanctions enforcement, and anti-money-laundering compliance, but labeled training examples are scarce because each one requires significant analyst effort to confirm. The idea of training two classifiers, one that looks at each actor on its own and one that also looks at the network around the actor, and using their disagreement as an active-learning signal goes back to Bilgic, Mihalkova, and Getoor (2010). C2AL extends that idea by combining the disagreement signal with an aggregated per-node uncertainty score from both classifiers, and the resulting sample-efficiency improvement holds across both real Bitcoin networks and synthetic anti-money-laundering benchmarks. On the hardest dataset in our suite, C2AL was the only method that reached the target accuracy within the labeling budget. These savings make automated screening of large transaction networks practical for institutions with limited investigation capacity.
Perspectives
Active learning is an important problem across many domains where labeled data is hard to come by. Relating the structure of the data, in our case the network structure of financial transactions, to the active learning process is a step in the right direction.
Amro Alabsi Aljundi
University of Virginia
I find the C2AL model exciting because it solves both technical and human problems. On the human side, analyst attention is a finite resource and it is demoralizing to stare down a huge list of data points that each require bespoke investigation- particularly when you know that the quantity of data in the system is growing rapidly and that the vast majority will not be of concern. By algorithmically targeting the highest-value nodes, C2AL can reduce human burden, thereby addressing a fundamental bottleneck that limits the development of models to improve the detection of financial bad actors.
Margaret Foster
University of Virginia
Read the Original
This page is a summary of: Network-based Active Learning for Identifying Illicit Actors in Financial Transaction Networks, International Foundation for Autonomous Agents and Multiagent Systems,
DOI: 10.65109/gock1684.
You can read the full text:
Contributors
The following have contributed to this page
