This article examines machine learning approaches used for fraud detection in loyalty programs, treating loyalty abuse as a distinct analytical problem rather than a simplified extension of payment fraud. The topic is timely because contemporary loyalty ecosystems combine account-based stored value, omnichannel interaction data, partner integrations, and promotion logic, which together generate heterogeneous fraud patterns and unstable labels. The article aims to systematize the prominent model families used in fraud analytics and determine which are most suitable for loyalty-program environments. The study relies on source analysis, comparative review, conceptual synthesis, and analytical generalization. Recent research on fraud analytics, anomaly detection, graph learning, tabular modeling, behavioral biometrics, explainable artificial intelligence, and adaptive risk estimation is examined. The analytical part identifies the strongest methodological trajectories for loyalty fraud detection, including hybrid pipelines, graph-based modeling, and behavior-aware scoring. The findings apply to program operators, fraud analysts, and product teams designing AI-supported decision systems for account protection, redemption control, and abuse prevention.