In particular, the CDW algorithm should attain higher accuracies for tasks
where feature importance varies according to the class. For example, CDW shows
high accuracy on Construct which is designed to exhibit varying feature importance.
Interestingly, CDW performs signicantly worse on all the NLP tasks.
We suspect that the important features for these tasks are unrelated to the class
of the instance. CDW may be basing its weights on spurious patterns in the
training data, lowering its accuracy