Explain unsupervised fraud and evaluation

Explain unsupervised approaches for fraud detection and when you would use them versus supervised methods. Compare options such as clustering, density estimation, isolation forests, autoencoders, and graph anomalies. Then discuss how to evaluate without reliable labels: use precision@k, recall at a fixed review budget, PR-AUC vs ROC-AUC under extreme imbalance, rank-based metrics, proxy/delayed labels, and calibration checks. Clarify why raw “accuracy” is misleading here and how you would choose thresholds.