A training approach combining focal loss (which emphasizes hard examples) with contrastive learning to handle imbalanced datasets.