Random forests are an ensemble method in supervised machine learning that extends bagging of decision trees by explicitly decorrelating the base learners. Because the variance reduction achieved by averaging an ensemble is maximal when its member predictors are uncorrelated, and bootstrap resampling alone leaves trees highly correlated (strong predictors dominate the top splits of every tree), random forests restrict each node's split search to a random subset of m of the p available features. This random feature subsampling at every node lowers inter-tree correlation and therefore yields a larger reduction in the variance of the averaged estimate, placing random forests alongside gradient-boosted trees as a leading tree-ensemble family.
Random forests are an ensemble method in supervised machine learning that extends bagging of decision trees by explicitly decorrelating the base learners. Because the variance reduction achieved by a…