Impact of the Recursive Feature Elimination. Impact of removing 10% of the features with the lowest weight vector in each round. After 30 iterations, with only 4.28% of all features remaining, the f-measure has dropped only by 2%. The underlying evaluation method only considers the recognition of single tokens rather than whole phrases. The bottom line (65 iterations) shows the impact of the remaining 0.11% of all features. All values are evaluated without the post-expansion step (see text).
Hakenberg et al. BMC Bioinformatics 2005 6(Suppl 1):S9 doi:10.1186/1471-2105-6-S1-S9