Regression trees pruning

Regression tree pruning reduces the risk of overfitting by verifying the predictive utility of all nodes of a regression tree. Nodes that do not improve the expected prediction quality on new data are replaces by leaves.

This decision is based on pruning criteria. To compare the predictive quality of nodes and leaves and to find out which ones are to be replaced, the reduced error pruning (REP) algorithm uses a separate pruning set.

The most commonly used quality indicator is the mean square error. You can, however, also use linear correlation or rank correlation as measures for regression quality.