A stopping criterion tells a recursive decision-tree algorithm when to stop splitting data. Without one, leaves can become tiny and the tree may overfit. A node can stop when its samples are already pure, or when a chosen limit prevents further subdivision. section 9.5
Common stopping criteria include:
- Stop when the improvement in accuracy, Gini index, or entropy falls below a threshold. section 9.5
- Stop when a node contains fewer than a specified number of samples. section 9.5
- Allow a split only when both resulting leaves contain enough samples. section 9.5
- Stop after the tree reaches a specified maximum depth. section 9.5
Stopping during tree construction
The algorithm repeatedly selects a leaf, finds its best feature split, and checks the stopping criteria. If splitting is allowed, it creates two new leaves; otherwise, the node remains a leaf and predicts its samples’ most common label. In an app-recommendation tree, traversal ends at such a leaf to recommend Atom Count, Beehive Count, or Check Mate Mate. section 9.6













