Fraud detection often has few labeled fraud cases and a much larger pool of unlabeled, mostly legitimate transactions. Semi-supervised learning uses both pools to reduce reliance on expensive manual labeling while exposing a model to more data patterns. section 7.4.3
Two common approaches use model predictions to make unlabeled data useful:
- Self-training: train an initial model on labeled data, add high-confidence predictions on unlabeled data to the labeled set, and retrain. section 7.4.3
- Co-training: use multiple models, often trained on different feature subsets, to label unlabeled instances where they have high agreement or confidence. section 7.4.3
These approaches can improve performance and generalization, but the benefit is not guaranteed. Synthetic data generation is a related approach when labeled fraud data is especially scarce or lacks diversity. section 7.4.3





