Classification - shivamvats/notes GitHub Wiki
- Don't use decision trees to predict class probabilities, the probabilities are biased towards 0 and 1. If you do have to do that, prune the tree first. It is better to use random forests for the job. Even then, the probabilities show a bias towards 0.2 and 0.9 (i.e., 0 and 1 are rare).