Plot the precision-recall curve

More informative than ROC when the classes are imbalanced, which they are here by construction: a 90th-percentile threshold makes only 10% of stations patches. ROC uses the false positive rate, whose denominator is the large non-patch class, so a model can look excellent while most of its positive predictions are still wrong. Precision asks the question that actually matters for a patch map — of the places called patch, how many are?

Usage

plot_pr_curve(predictions, path = NULL)

Arguments

predictions

the predictions element of a fit_patch_model() result

path

where to write a PNG; NULL returns the plot instead

Value

the plot object, or path invisibly when written to disk

Details

The baseline is the patch prevalence: what precision random guessing achieves.