A fraud-detection model that predicts "not fraud" every single time can still score 99% accuracy if fraud is rare in the dataset — and it would be completely useless. This is exactly why accuracy alone can't be trusted for many real classification problems.
Precision
Precision answers: of everything the model flagged as positive, how much was actually correct? High precision matters when false positives are costly — flagging a legitimate transaction as fraud, for instance, frustrates real customers.
Recall
Recall answers: of everything that was actually positive, how much did the model catch? High recall matters when missing a positive case is the costly mistake — failing to catch an actual fraudulent transaction, for example.
Most real projects require a deliberate trade-off between the two, tuned to the actual cost of each type of mistake. Our Advanced Machine Learning course spends real time on this judgment call, not just the formulas behind it.
