MAS-II: Tree pruning separates training fit from selection
A deep tree can closely fit a training sample and still predict poorly. Pruning or controlling complexity changes the bias-variance trade-off. Evaluate the whole procedure, including tuning, on data it did not use to select the tree.
Worked example or practice scenario
A maximal tree has training error 1 and validation error 18; a pruned tree has training error 5 and validation error 10. The apparent training deterioration can accompany a useful predictive improvement. The split sample and metric must be the same for a fair comparison.
Try this next
Describe how cross-validation would choose the complexity parameter. Then test sensitivity to a different split. Do not treat one lower validation score as proof that the selected structure is stable.
Reading sources
ActNet editorial guide · October 1, 2026 · Original illustrative examples.