- only datasets with more than 1000 instances
- TwoStageCollective with only one iteration, i.e., no flipping
- VotedFolds used, to speed up computation for algorithms that are quadratic
or worse in regards to number of instances
- three different setups were used
Split 5/95%, VotedFolds with 19 folds
Split 10/90%, VotedFolds with 9 folds
Split 20/80%, VotedFolds with 4 folds
The collective classifier often performs better with more folds, e.g., SMO
either wins or has a tie with 19 folds. With fewer folds it starts to lose
against the base classifier.
J48, on the other hand, seems to perfom better with fewer folds. It manages to
win against the base classifier on the waveform dataset.