In a classification forest, strength is the expected margin across observations: how decisively the forest favors the known class over its strongest competitor. section 9.4.2
The equation defines strength s as the population expectation of the forest margin. section 9.4.2
Estimate strength by averaging margins across labeled out-of-bag or test observations. A large positive value signals a consistent voting advantage; a value near zero signals divided votes, and a negative value signals poor classification. section 9.4.2
The generalization-error bound relates forest error to strength s and average tree correlation ρ̄: greater strength helps, while greater correlation weakens the bound. section 9.4.2
Strength is necessary but not sufficient for useful aggregation: strong, highly correlated trees repeat errors, while diverse but weak trees contribute little reliable information. Random feature selection can reduce correlation without sacrificing too much strength. section 9.4 section 9.4.2
