AI and technology

Semi-Supervised Machine Learning

A hybrid machine learning approach that trains models on a small set of labeled data combined with a much larger set of unlabeled data. It is useful in AML, where confirmed money laundering cases are scarce compared with the volume of unlabeled transactions.

Why it matters

Labels in financial crime are expensive and incomplete. Only a small share of alerts lead to a suspicious activity report, and investigators rarely learn whether a case ended in a conviction. Supervised models trained on so few examples tend to overfit to past cases, while unsupervised models can surface activity that is unusual but not suspicious. Semi-supervised methods try to combine the strengths of both.

Common techniques include self-training, where a model labels the unlabeled records it is most confident about and retrains on them, and approaches that first learn the structure of normal behavior from unlabeled data, then use the few labeled cases to calibrate what counts as risk. Investigator decisions on alerts become a steady source of new labels.

Supervisors expect models used for monitoring to be explainable, validated and governed under model risk management standards. Teams should document how labels were chosen, how machine-generated labels are controlled, and how the model performs on new typologies as well as known ones.

Protect the flow. Power the growth.

See how ThetaRay turns regulatory expectations into detection and investigation that hold up to scrutiny.