##### Scenario You receive millions of historical transactions but without fraud labels. Management wants an unsupervised system to surface potentially fraudulent transactions and a way to evaluate its effectiveness. ##### Question Which unsupervised learning approaches would you choose to flag suspicious transactions and why? Without labels, how would you measure the accuracy or performance of your model? Name concrete evaluation techniques or proxy metrics. ##### Hints Think clustering, distance-based anomaly detection, autoencoders, human review samples, precision from post-labeling, business KPIs.