Figure 1

Classification error. Estimated Out-of-Bag (OOB) Error with regard to progressive cancer class removal. At each step, cancer classes contribute most to classification error were removed and OOB errors were recalculated. B: Breast Cancer. P: Prostate Cancer. O: Ovarian Cancer. C: CNS. The left plot shows OOB errors on entire dataset and the right plot shows OOB errors on subset of classifiers selected by varSelRF. The improvement of variable selection is more remarkable for microarray because of the much larger classifier pool to select from (11961 v.s. 342). Metabolite classifiers can achieve average OOB error of 0.28 when B,P,O,C are removed, reduced from 0.51 from the full set. The OOB error for u133a reduced from 0.34 to 0.18. The average OOB errors from the random cancer class removal are plotted in the orange dashed line.