Metrics coverageerror

CoverageError

How far down the ranked labels you have to read before you have seen every relevant one, averaged over the samples. A sample whose two relevant labels are the two highest-scoring covers 2; the same sample with one of them scored last covers the whole row.

The floor is not 1 but the number of relevant labels the row holds — scikit-learn's own "The best value is equal to the average number of labels in y_true per sample." A coverage of 2.5 is perfect on samples carrying two or three relevant labels and poor on samples carrying one, so the number means nothing without the label counts beside it.

A sample with no relevant label contributes 0, not the label count. There is nothing to cover, and the row is not dropped from the average either, so the mean can sit below 1 — measured, two samples one of which is empty give 0.5. Read a coverage under 1 as "some rows had nothing to find", never as an impossibly good ranking.

Like LabelRankingLoss and unlike LabelRankingAveragePrecision, a single label column is refused with scikit-learn's sentence, binary format is not supported. That the three do not agree about it is a divergence inside the reference, reproduced rather than smoothed.

Members

Member	What it does
CoverageError.Score	The mean rank of the worst-ranked relevant label.