Metrics labelrankingaverageprecision

LabelRankingAveragePrecision

The average-precision idea applied to one sample's labels. For each relevant label, ask what fraction of the labels ranked at or above it are themselves relevant; average that over the relevant labels, then over the samples. 1 means every relevant label sits above every irrelevant one, and the number degrades gracefully as irrelevant labels work their way up.

Where CoverageError reports a position and LabelRankingLoss reports a fraction of lost comparisons, this reports how clean the top of each ranking is, and it is the one of the three that is scale-free and bounded without knowing how many labels are relevant.

It accepts a single label column and returns 1, where the other two refuse it with binary format is not supported. Nothing here decides that: label_ranking_average_precision_score validates its input differently from coverage_error and label_ranking_loss, and the divergence is reproduced rather than smoothed — making the three agree would invent a difference from the reference instead of copying one.

A weight vector summing to zero returns NaN here and raises in the other two, for the same kind of reason: the reference divides by the weight sum directly on this path and calls numpy.average on the other two, and only numpy.average refuses a zero sum.

A sample where every label or no label is relevant scores 1 — its ranking carries no information, and the reference says as much in a comment of its own.

Members

Member	What it does
LabelRankingAveragePrecision.Score	The mean, over relevant labels, of how much of the ranking above them is relevant.