Predicting Error Rates
Suppose it was possible to predict the error rate of a tree and of its subtrees (including leaves)
Then, pruning rationale:
- Start from bottom of the tree and examine each nonleaf subtree
- If replacement of this subtree with a leaf, or with its most frequently used branch, would lead to a lower predicted error rate, then prune
- Since the error rate for the whole tree decreases as the error rate of any of its subtrees is reduced, this process will lead to a tree whose predicted error rate is minimal