F. Model performance
Figure 5 shows the decision tree model built from our oracle. The root node of the tree is ChangeTrigger 1 (i.e. changes within one line distance to a comment-highlighted line).
Comment status is the second most important attribute for the classification. These two most important signals were evident to us during the first stage of this study (Section IV-B), and our model provides confirmation. The other useful attributes for the model are number of comments, iteration number, and sentiment group.
Based on one hundred 10-fold cross-validations of the decision tree model, our model had a mean precision of 89.1%, mean recall of 85.1%, and mean classification error of 16.6%. Finally, the review participants that we contacted rated 97 out of the 125 comments as Useful. Our model classified 105 of those review comments as ‘Useful’, of which 91 were correct. In this step, our model had 86.7% precision, and 93.8% recall.