Correlation describes the strength and direction of association between quantities. In a random forest, it helps explain whether combining many trees produces genuinely different predictions or mostly repeats the same information. section 9.4.2
For classification forests, correlation concerns how trees’ margin behavior moves together across observations. It is high when trees tend to classify and misclassify the same observations; it is lower when their successes and errors occur on different observations. Highly correlated trees offer limited gains from voting, even when each tree is strong. section 9.4.2
Diversity alone is not enough: weak, nearly random trees may make different errors but contribute too little reliable information. A useful forest balances low correlation with strength, defined as the expected margin—the voting advantage of the known class over its strongest competitor. section 9.4.2
This upper bound relates classification generalization error to forest strength s and average correlation ρ̄. It is a bound, not an exact prediction of error. section 9.4.2
Random feature selection can reduce correlation by preventing the same dominant predictors from shaping every tree. Considering fewer features can also weaken individual trees, so the trade-off matters. Adding trees stabilizes the ensemble’s prediction, but cannot by itself repair weak or highly correlated learners. section 9.9 section 9.4.3
Correlations under changing conditions
A pedestrian-detection network trained mostly on clear daytime images may rely on background or lighting correlations that reduce training loss but fail at night or in severe weather. More daytime images do not fix the missing coverage; performance should be monitored as operating conditions change. section 11.6.3





















