AUPRC

Classification metric computing the per-head macro-averaged area under the precision-recall curve. Each class is scored one-vs-rest: sweeping a threshold down over that class’s predicted probability traces precision against recall, and the area under that curve – average precision – summarizes ranking quality across every operating point rather than at the single threshold an arg-max implies. The per-class areas are then averaged with equal weight. Unlike AUROC, the baseline is the class prevalence, which makes AUPRC the more honest read on a rare signal class.

Classes with no positive samples in the evaluation set are left out of the average.

Approximation

The exact curve needs every score held and sorted, which an incremental accumulator cannot do. Scores instead go into bins equal-width bins on [0, 1], kept separately for positives and negatives, and the threshold sweep runs over bin edges. The area is exact to bin resolution, with ties inside a bin resolved as if positives and negatives interleaved evenly. Raising bins tightens the estimate at linear cost in a small histogram, and the bin count is reported as part of the metric name (ie auprc100).

Configuration

Selected as name: auprc.

Option	Description	Type	Default
bins	Number of equal-width score bins on [0, 1]; sets the resolution of the threshold sweep.	int	100
from_logits	Treat the model output as unnormalized logits. Set to false when the model already emits log-probabilities.	bool	true

services:
  metrics:
    select:
      - name: auprc
        kwargs: { bins: 200 }