Figure 2.

Performance of the unified generative-discriminative learning principle for four data sets. We perform a 1,000-fold stratified hold-out sampling procedure for the four data sets, record for different values of β the mean sensitivity for a fixed specificity of 99.9%, and plot the mean sensitivities on the simplex β in analogy to Figure 1. Yellow indicates the highest sensitivity, red indicates the lowest sensitivity, and the gray contour lines of each subfigure indicate multiples of the standard error of the maximum sensitivity.