Precision-recall curve

Plots precision against recall for a binary classifier and reports the average precision (the area under the curve). The precision-recall curve is more informative than the ROC curve when the positive class is rare, because it ignores the many true negatives. A horizontal line marks the no-skill baseline (the positive-class prevalence). Pass a named list of models / (actual, score) pairs to overlay several colour-coded curves with a legend and per-curve average precision.

Usage

pr_curve_plot(
  x,
  score = NULL,
  colour = depictr_brand(),
  f1 = FALSE,
  title = NULL
)

Arguments

x

A binomial glm; the vector of observed outcomes (0/1, logical or a two-level factor with the positive class second); or a named list of models / (actual, score) pairs to overlay.

score

When x is an outcome vector, the matching scores or predicted probabilities (or a list of them for the multi-model case).

colour

Curve colour for the single-model case. Defaults to the depictr brand blue. Ignored when several models are overlaid.

f1

Logical; if TRUE, mark the maximum-F1 operating point on each curve (the threshold maximising the harmonic mean of precision and recall).

title

Plot title.

Value

A ggplot2::ggplot object. The average precision is stored in attr(plot, "average_precision") (a named vector when several models are supplied).

Examples

# The clinical-trial adverse event is rare (about 10% positive), which is
# where a precision-recall curve is worth drawing.
gfit <- glm(adverse_event ~ biomarker + age + arm,
            data = clinical_trial, family = binomial)
pr_curve_plot(gfit)


# Mark the maximum-F1 operating point.
pr_curve_plot(gfit, f1 = TRUE)


# Compare two models with per-curve average precision in the legend.
reduced <- glm(adverse_event ~ biomarker, data = clinical_trial,
               family = binomial)
pr_curve_plot(list(Full = gfit, Reduced = reduced))