1
$\begingroup$

I think I understand what pruning is (in concept) in a decision tree. What is not so clear to me (in reading) is what happens to the pruned observations. What do you do with them? Are they just simply not used and forgotten?

What bothers me is the "hole" left behind by the pruned node. What if a future $Y$ needs to be predicted at that node? Do you go to the parent node and treat that parent node as an end node and predict $Y$ by the data at that parent node (even though some data at the parent node continue to be divided in sub-nodes, other than the pruned node)?

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ What do you mean by incidents here exactly? $\endgroup$
    – Valentin Calomme
    Commented Jan 6, 2021 at 7:28
  • $\begingroup$ I thought Incidents meant "observations" in ML terminology. Is that not true? $\endgroup$
    – Student of Statistics
    Commented Jan 6, 2021 at 14:49
  • 1
    $\begingroup$ I personally never heard that term. Typically, I would use terms like inputs, data points, input features/variables, instances. And what you're trying to predict is generally referred to as the target, gold truth, label. $\endgroup$
    – Valentin Calomme
    Commented Jan 6, 2021 at 15:28

2 Answers 2

Reset to default
1
$\begingroup$

Once you prune nodes, they get removed from the tree altogether, so you wouldn't be able to predict based on them.

Typically, hyperparaneter tuning is used to ensure that the right nodes are pruned. Obviously, there is always a trade-off.

Improve this answer
$\endgroup$
0
$\begingroup$

Pruning doesn't operate on individual leaves, it happens to splits: you decide not to make a split at a node after all, and so all the observations that would go through that node in the original tree will stay at that node, which is now a leaf.

Improve this answer
$\endgroup$
2

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.