A fraud dataset is highly imbalanced and a model predicts every transaction as legitimate while reporting high accuracy. Which approach best fits?

1. Investigate overfitting and use appropriate regularization or data improvements.
2. Use metrics such as precision, recall, and PR analysis appropriate to the business costs.
3. Choose and validate a decision threshold aligned with recall and false-positive costs.
4. Reserve a final independent test set and use validation data for model selection.