Figure 4.

Swi6 Feature Ranking. A feature ranking is created on every training set during cross validation of a classifier. Since 50 classifiers are made for every TF, this results in hundreds of separate rankings (50 * #Training Points). This plot shows the frequency with which a feature shows up in the group of top 40 features. When sorted by this frequency, only the most important features remain at the top of the list. For Swi6 the first 10 features are in the top-40 of 60% of the rankings, making them consistently chosen over many resamplings of the negative set.