Metrics 0.3.0 precisionrecallcurve

Lodestar.Metrics 0.3.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

PrecisionRecallCurve

The precision-recall curve as plot data. Where RocCurve barely moves when a few thousand negatives are ranked above a handful of positives, this collapses — which is why it is the curve to plot when positives are rare.

Its thresholds array is one shorter, deliberately

Precision and Recall have one more entry than Thresholds. The extra point is the endpoint at recall 0 and precision 1, which no threshold produces — it is where the model predicts nothing positive at all.

Padding the array to match would invent a threshold for that point, and a caller plotting thresholds against precision would silently plot one pair too many. The asymmetry is the reference's, and decision 0040 keeps it.

The area under it is not the average precision

Auc.Trapezoid over these points interpolates between two thresholds as though the curve were straight there, and reads optimistic: AveragePrecision.Score sums the steps instead. Measured on the worked case, 0.7916… against 0.8333…. A test holds the two apart rather than a reader having to.

dropIntermediate is false here, as the reference has it, and drops by a different rule from RocCurve's — see that page's table.

Members

Member	What it does
PrecisionRecallCurve.Compute	Draws the curve from labels and scores.