The Tree That Grew Too Deep

A decision tree grown with no depth limit or minimum-leaf-size constraint achieves 100% training accuracy but generalizes poorly to new data. Which technique fixes this by growing the full tree first and then removing branches that don't improve performance on held-out data?

Increasing the maximum tree depth so the model can separate the classes more cleanly

Post-pruning (e.g., cost-complexity pruning), which trims back a fully grown tree using validation performance

Setting a minimum number of samples required per leaf before the tree is grown, so small, noise-fitting splits are never made

Standardizing the input features before training