When building decision trees over a dataset that generates nodes with bad purity, is there any benefit of using the CART algorithm over the iterative dichotomizer 3 (ID3) algorithm?
CART does binary splits. ID3, C45 and the family exhaust one attribute once it is used. This makes sometimes a difference which means that in CART the decisions on how to split values based on an attribute are delayed. Which means that there are pretty good chances that a CART might catch better splits than C45.
The drawback is that with CART you can't create rules and the whole tree is larger and harder to interpret. Anyway, the interpretation is not always useful.