Given a highly imbalanced dataset where fraud represents less than 1% of the samples, which evaluation metrics would you prioritize and why?
