Compute CU only in terms of the instance attributes, this should help minimize overfitting and prevent arbitrary predictions about missing attributes that arise from it being better to recurse on a node that has no attributes in common than to create a new node.
Compute CU only in terms of the instance attributes, this should help minimize overfitting and prevent arbitrary predictions about missing attributes that arise from it being better to recurse on a node that has no attributes in common than to create a new node.