Comparison of learning curves from models that predict whether PWUDs would use meth within the next 12 months using (a, b) small and (c, d) large number of features.

<p>Plots (a) and (b) correspond to the LG and DT models with features selected from the top-<i>k</i> correlated method that returned the highest scores (the same model yields the highest AUROC and AUPR for both classifiers of meth) in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0312046#pone.0312046.t004" target="_blank">Table 4</a>, which include 5 and 4 features, respectively. Plots (c) and (d) correspond to the LG and DT models that incorporate 165 features from the largest manually-grouped combination (SCDM+ SUB+ TREX). Each plot displays training (red) and cross-validation (green) AUROC as function of number of training examples. The dashed line marks the baseline (all 0.5 for AUROC).</p> <p>(PDF)</p>