For a labeled observation, a forest margin is the forest’s voting support for the correct class minus its support for the strongest competing class. It measures voting decisiveness, not just whether the winning class is correct; unlike a support vector machine’s geometric margin, it concerns collective votes. section 9.4
Forest strength is the expected margin across the population; labeled out-of-bag or test observations can estimate it by averaging their margins. section 9.4.2
Interpret an average margin as follows:
- A large positive value indicates a consistent voting advantage for the known class. section 9.4.2
- A value near zero indicates frequently divided votes; a negative value indicates systematically poor classification. section 9.4.2
Margin-based strength is only part of the generalization story: useful forests need trees that predict better than chance and do not repeatedly make the same errors. Breiman’s error bound relates forest generalization error to strength and average tree correlation. section 9.4 section 9.4.2
