Metrics 0.3.0 averageprecision perlabel

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.

AveragePrecision.PerLabel

One average precision per label of a matrix, uncombined — average_precision_score(…, average=None).

public static double[] PerLabel(ReadOnlySpan<bool> yTrue, ReadOnlySpan<double> yScore, int labelCount, ReadOnlySpan<double> sampleWeight = default)

Parameters — yTrue says whether each label is relevant and yScore holds the scores, both row-major: one row per sample, labelCount values each, and the same length. labelCount is how many labels a row holds. sampleWeight is one weight per sample — per row, not per label — or empty, the default.

Returns — double[], one score per label in column order, each of them what AveragePrecision.Score's binary overload would return for that column on its own. A label no sample carries scores 0 rather than being dropped, which is why the Macro mean over these is not the mean over the labels that actually occur.

Exceptions — ArgumentException when labelCount is below 1, when yTrue and yScore disagree in length, when yTrue is empty or not a whole number of rows of labelCount, or when a non-empty sampleWeight is not one per row.

Example — two samples over three labels, where the middle label is carried by neither.

using Lodestar.Metrics;

bool[] relevant = [true, false, false, false, false, true];
double[] labelScores = [0.75, 0.5, 1.0, 1.0, 0.2, 0.1];

double[] perLabel = AveragePrecision.PerLabel(relevant, labelScores, labelCount: 3);
double middle = perLabel[1];  // => 0

The outer two score 0.5 each and the middle 0, which is the 0.3333… that Averaging.Macro reports and the 0.5 that Averaging.Weighted reports — the two differ only in whether the empty column is counted.

Applies to — net10.0, netstandard2.0.

See also — AveragePrecision.Score, the Python equivalence table.