Split data into training and
Do until further pruning is harmful:
- Evaluate impact on
validation set of pruning each possible node (plus those
- Greedily remove the one that most improves
validation set accuracy
- produces smallest version of most accurate subtree
- What if data is limited?