Fig 6.

The performance of four feature selection methods in subset pair ‘Prostate cancer versus Normal’ of GSE71008 using linear SVM with different numbers of features. All four feature-selection methods had a similar performance ranging from 1 to 50 features. Ridge regression and linear SVC performed poorly when less than 5 features were selected (AUCs < 0.80), while the random forest method (0.85 < AUCs < 0.90) and lasso regression method (0.75 < AUCs < 0.95) performed pretty well. The random forest method performed better when fewer features were selected (14–22 features, AUCs > 0.95), while other methods had increasing AUCs (AUCs > 0.80) when more features were selected