Stopping Conditions

Learn when decision trees stop splitting and how to control overfitting.

Without limits, a decision tree can keep splitting until every leaf is pure—or until each leaf holds a single sample. That leads to overfitting.

Big idea

Stopping conditions tell the tree when to stop growing and make a leaf instead of splitting again. A node becomes a leaf when any of these is true:

All samples share the same class (pure node: Gini $= 0$ or entropy $= 0$ )
A hyperparameter limit is reached
No split improves impurity enough

After the root split, the tree recurses only on child groups that are still impure and have not hit a stopping rule.

Formula

A split is worth keeping when impurity drops:

\text{Impurity(parent)} > \text{Weighted impurity(children)}

If no candidate split lowers weighted impurity (or the improvement is below a threshold), the node becomes a leaf.

Common scikit-learn controls:

Parameter	What it does
`max_depth`	Maximum number of split levels from the root
`min_samples_split`	Minimum samples required to split a node
`min_samples_leaf`	Minimum samples each leaf must hold
`min_impurity_decrease`	Minimum impurity reduction a split must achieve

Worked example

On our practice tree, Feature 0 $\leq 2.5$ creates a right child that is pure (all class 1)—that branch stops immediately.

The left child is still mixed, so the tree recurses. Eventually Feature 1 at $t = 0.5$ splits the last impure group into two pure leaves.

If we set max_depth=1, the tree would stop after the root split—even though the left child is still mixed. Tuning these knobs balances underfitting (too shallow) and overfitting (too deep).

Common mistake

Setting max_depth=None with no other limits on a small dataset. The tree may memorize every training sample, including noise, and fail on new data.

Key takeaway

Stopping conditions prevent trees from growing forever. Use max_depth, min_samples_split, min_samples_leaf, and min_impurity_decrease to control model complexity.

Experiment with these parameters in the practice notebook below.

Practice

Try It Yourself

Open the practice lab to complete the starter code in the notebook.

Knowledge Check

Quick Quiz

Which parameter limits how deep a tree can grow?

Summary

Key Takeaways

Trees stop when nodes are pure or reach a depth limit.
max_depth, min_samples_split, and min_samples_leaf control growth.
Stopping early helps prevent overfitting.