The experiment results of all the combination strategies and all the participating single classifiers
are shown altogether in Figure 8.
From Figure 8, we have the following conclusions and discussions:
- Higher feature dimension generally leads to better performance (the growth slows
down as the feature dimension increases).
- Impurity measure generally works better than posterior variance for feature selection.
- Classifier combination methods do not necessarily work better than single classifiers.
- Majority vote combination strategy works the best. This actually surprises us a little
bit since we originally expected the dynamic classifier combination strategy to produce
the best result.
