How does Random Forest behave under label noise with symmetric vs asymmetric corruption?

Random Forest is robust to symmetric noise but significantly more sensitive to asymmetric label noise.

In symmetric noise, labels are randomly flipped uniformly, which averages out across trees. In asymmetric noise, specific classes are systematically mislabeled, causing biased splits and reinforcing incorrect decision boundaries. Since tree splitting depends on impurity reduction, systematic label corruption directly distorts tree structure and propagates through the ensemble.