How does Random Forest behave under feature collinearity at extreme scale?

High collinearity reduces tree diversity and weakens variance reduction benefits.

When features are highly collinear, multiple features carry redundant predictive signal. Random Forest’s feature subsampling may repeatedly select correlated features, leading to similar splits across trees. This increases inter-tree correlation (ρ), reducing ensemble variance reduction efficiency. At extreme scale, RF behaves closer to bagged trees with limited decorrelation benefit.