Improving inductive accuracy by classifying errors