Figure 6

Per-species ensemble of classification results. (A–E) Receiver operating characteristic (ROC) curves representing the true positive rate (TPR) as a function of the false positive rate (FPR) across all model thresholds of classification, labeled by method. All non-RNN classification methods are conducted using population references. (F) Table of per-species precision and recall across all surveyed methods (N=100). Bold statistics in the table represent the highest performer for each metric and species. Precision values for P. bethaniensis and P. forresti are low because these two species represent the classes with the fewest number of samples, and so there is a very small amount of true positive values. However, these still greatly exceed what would be expected by chance (0.001 and 0.006, respectively). The high recall for these classes indicates that the true positives are correctly captured.