Which RandomForestClassifier hyperparameter most directly controls the tradeoff between individual tree diversity and individual tree accuracy?
Select an answer to reveal the explanation.
Short Explanation and Infographic
featureSubsetStrategy is the diversity dial — restrict each tree to a random subset of features at each split, and trees disagree with each other more, which is exactly what makes ensemble averaging powerful.
Full explanation below image
Full Explanation
featureSubsetStrategy in RandomForestClassifier controls the number of features sampled at each tree node split. Options: 'auto' (default, = sqrt(numFeatures) for classifiers, numFeatures/3 for regression), 'sqrt', 'log2', 'onethird', 'all', or a fraction/integer. Why it controls bias-variance: small featureSubsetStrategy → trees see fewer features → more diverse, less correlated trees → better ensemble variance reduction (but each tree has higher bias). large featureSubsetStrategy (approaching 'all') → trees are similar, highly correlated → less ensemble benefit, each tree has lower bias but high variance. Random Forest theory (Breiman 2001) shows that decorrelating trees via random feature selection is the key to the ensemble's generalization. numTrees reduces variance but has diminishing returns. subsamplingRate adds row-level randomness (row bagging). maxBins affects how continuous features are split (finer splits with more bins).