Figure 4

Comparison of feature selection approaches in terms of predictive accuracy. The prediction performances of the models learned by greedy RLS were assessed using area under the ROC curve (AUC), averaged over the three folds of an external CV. On each round of the external CV, the training set on which the features are selected consisted of 2/3, and the independent test set on which the prediction performance is measured, 1/3 of the 3,410subjects, with a stratified training/test split. The graph also displays the individual SNP AUCs for each of the variants selected by greedy RLS. Further, results are depicted for a p-value based filtering in which the top k most significant features were selected. We also present a curve that displays the results for the hybrid method in which greedy RLS runs on the top k features ranked according to their p-values. Finally, we present a random permutation on the class labels and running greedy RLS on this randomized dataset.