Figure 2

Subfigure 2(a): Synthetic data. Subfigure 2(b): Colon cancer. Comparison of leave-one-out error curves for synthetic data and real data sets. The model error is computed on the data previously used for ranking. On synthetic data (a), the RFE-SVM method achieves perfect classification with 9 of the 1000 relevant features (solid line) in data set f1000–5000. Moreover, 20 features are sufficient to reach perfect classification on the purely noisy data set f0–5000 (dashed-dotted line). In the right panel (b) for the Colon cancer data from ref. [23], similar error estimates are obtained for the real data (perfect classification with 8 genes – solid curve) and with randomized labels (dashed curve), for which 14 genes are sufficient to get a zero error estimate.