Per-Class Recall

Classification metric computing the per-head, per-class recall: for each class, the fraction of samples truly in that class that were predicted as it. This metric reports one value per class.

It is the breakdown the macro averages summarize, and the place a collapsed minority class shows up. A class the evaluation set never contains has undefined recall and is reported as nan.

Configuration

Selected as name: per-class-recall. Takes no options.

services:
  metrics:
    select:
      - name: per-class-recall
        kwargs: {}