Metrics 0.3.0 precisionrecallcurve compute

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.Compute

Draws the precision-recall curve — sklearn.metrics.precision_recall_curve.

public static PrecisionRecallCurve Compute(ReadOnlySpan<int> yTrue, ReadOnlySpan<double> yScore, int posLabel = 1, ReadOnlySpan<double> sampleWeight = default, bool dropIntermediate = false)

Parameters — yTrue is the true labels, one per sample. yScore is a score per sample. posLabel is the label counted as positive, 1 by default. sampleWeight is one weight per sample, or empty. dropIntermediate drops points whose true-positive count matches both neighbours; false here, as the reference has it.

Returns — a PrecisionRecallCurve whose Precision and Recall are the same length and whose Thresholds is one shorter, for the reason the type page gives.

Exceptions — ArgumentException when the inputs disagree in length, are empty, or hold a NaN score.

Example — the asymmetry, in one line.

using Lodestar.Metrics;

int[] truth = [0, 0, 1, 1];
double[] scores = [0.1, 0.4, 0.35, 0.8];

PrecisionRecallCurve curve = PrecisionRecallCurve.Compute(truth, scores);
int missing = curve.Precision.Count - curve.Thresholds.Count;  // => 1

Remarks — with no positive sample the recall is taken as 1 at every threshold, which is what the reference warns about and returns; AveragePrecision.Score reproduces the same substitution and its page says so.

Applies to — net10.0, netstandard2.0.

See also — RocCurve.Compute, DetCurve.Compute, AveragePrecision.Score, the Python equivalence table.