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