Which statements about decision tree pruning are true?

In decision tree pruning, the primary goal is to improve the model's predictive performance by reducing overfitting. One common issue with decision trees is that as they grow and make more splits, they can capture noise in the training data rather than the underlying pattern, which leads to overfitting. This is why the first statement is true; the recursive binary splitting method does indeed lead to overfitting, especially in situations with complex datasets where the model can become excessively detailed.

A tree with more splits typically increases the complexity of the model, which in turn can increase variance rather than decrease it. This is because more splits allow the model to fit the noise in the training data, which makes it less generalizable to unseen data. Hence, the second statement is not true.

Regarding the cost complexity pruning method, this technique employs a parameter (α) to control the trade-off between tree size and training accuracy. When α=0, the penalty for complexity is not applied, and as a result, the tree would be unpruned and retain its full size, which is consistent with the third statement being true.

Therefore, the combination of the first and third statements being true validates the choice that only those statements are correct.