Statistics and Actuarial Science
Simon Fraser University
Adaptively Pruned Random Forests Using Likelihood-Based Trees
Tuning the tree sizes in a random forest is not generally recommended, but we have found cases where the default node sizes are not adequate. However, tuning node size can be ineffective and crossvalidation for pruning trees within the forest is expensive. We develop a fast pruning method based on local likelihoods and a custom-developed information criterion (IC). The amount of pruning is controlled by adaptively adjusting the weight on the IC penalty. The method can automatically select the pruning level in roughly the same time as it takes to build the forest. In 13 example data sets, RMSE is never significantly increased and sometimes significantly lowered by using this pruning.
Refreshments at 3:45pm in Snedecor 2101.