Use macro average in confusion matrix metrics to normalize imbalanced classes

ConfusionMatrix multi class measures should be based on macro average to avoid bias (for the good or the bad) from imbalanced classes.