Comparison of TSP, Fisher and RFE as feature selection methods for SVM as correlation varies among signal genes. A) shows the error rates of SVM (mean ± SE) on the test set of Data-I (the single block structure) at various gene selection levels, as within-block correlation (ρ) varies. B) shows the error rates of SVM on the test set of Data-II (the multi-block structure), as inter-block correlation (ρ') varies. The horizontal lines are the error rates of SVM using all features. The two vertical lines in A) show the number of pairs of genes in which recovery of signal genes are examined as shown in Figure 3.