Unsupervised learning

Unsupervised learning finds structure in data that has no labeled outcome to predict. Instead of learning “default or no default,” it groups similar customers, places cases near the ones most like them, or flags records that do not fit the pattern the rest of the data follows.

Where it earns its place

Customer and portfolio segmentation, where the institution wants groups it can act on rather than one score. Anomaly detection in payments, claims, or sensor readings, where confirmed fraud or failure labels are rare or arrive late. Similarity and representation, where a case needs to be compared with the cases most like it. It often sits in front of a supervised model: segments become inputs, and anomalies become cases a person reviews.

How it is judged

There is no held-out outcome to score it against, so accuracy is the wrong question. A segmentation is judged on whether its groups stay stable across time windows and samples, whether each group means something the business can act on, and whether the result survives a change of method. An anomaly model is judged on the alert volume a team can actually review and on what reviewers find in those alerts. Both are monitored for drift like any other production model.

Where it sits in a larger system

A segment or an anomaly flag is an input, not a decision. Policy, limits, and the responsible reviewer decide what happens to a flagged case. The machine-learning models page covers segmentation and anomaly models alongside supervised ones, with the same lifecycle, documentation, and handover.