How does pruning reduce overfitting in Decision Trees?

Pruning removes unnecessary branches that do not improve generalization, reducing model complexity.

Pruning works by penalizing overly complex trees. In cost-complexity pruning (used in CART), a penalty term α * number_of_leaves is added to the loss function. The algorithm finds subtrees that minimize both impurity and complexity. This prevents the model from fitting noise in training data and improves generalization performance.