Figure_3.tif (562.82 kB)

Download file# Average accuracy of SVM, as a function of number of features.

figure

posted on 2013-12-02, 03:28 authored by Meysam Bastani, Larissa Vos, Nasimeh Asgarian, Jean Deschenes, Kathryn Graham, John Mackey, Russell GreinerFor each r = 1,2,…,18, line 3 of FS_SVM (Figure 2) computes the mean *a _{r}* and standard deviation

*σ*of the empirical accuracies obtained, over all 10 folds; this figure plots these bars, for each r. Notice the average accuracy on the hold-out sets increases as the number of features is increased, then levels out, with only minor fluctuations. Here, the largest accuracy occurs at r = 4; notice however that this accuracy is “essentially” the same as at r = 3. We therefore set r

_{r}^{*}= 3 as it is the smallest number of features whose accuracy's “mean + standard deviation” is at least the high-water-mark mean accuracy.