The guide
Accuracy is close to useless and almost every real problem shows you why. If one class is 99 percent of the data, predicting that class every time scores 99 percent while detecting nothing at all.
So the metrics here are about being honest under imbalance. Precision and recall pull in opposite directions and F1 is one way to trade them off. ROC AUC asks a different question entirely: over every possible threshold, how often does the model rank a positive above a negative? That makes it threshold-free, which is useful and occasionally misleading.
Macro and micro averaging exist because multi-class problems make you say whether you care about every class equally, or every example equally. They are different questions and the numbers can disagree sharply.