Random Forests provide information about the underlying dataset that most of other methods cannot easily provide. A prominent example is the importance of each individual feature in the dataset. One method to estimate feature importance is to use the Gini index for each node of each tree and compare each feature's cumulative value. Another method uses the out-of-bag samples. First, the out-of-bag accuracy is recorded for all base learners. Then, a single feature is chosen and its values are shuffled in the out-of-bag samples. This results in out-of-bag sample sets with the same statistical properties as the original sets, but any predictive power that the chosen feature might have is removed (as there is now zero correlation between the selected feature's values and the target). The difference in accuracy between the original and the partially random dataset is used as measure for the selected feature's importance.

Concerning bias and variance, although random forests seem to cope well with both, they are certainly not immune. Bias can appear when the available features are great in number, but only few are correlated to the target. When using the recommended number of features to consider at each split (for example, the square root of the number of total features), the probability that a relevant feature will be selected can be small. The following graph shows the probability that at least one relevant feature will be selected, as a function of relevant and irrelevant features (when the square root of the number of total features is considered at each split):