Close the boundary around one exact imbalanced table.
Keep the boundary
This book pins one imbalanced table and one always-N baseline.
one pinned table; one pinned baseline \text{one pinned table; one pinned baseline} one pinned table; one pinned baseline Class Imbalance Exactly Exact row counts for one always-N baseline. label counts label count N 4 P 1 always-N baseline on one imbalanced table row true pred correct? cell A N N yes TN B N N yes TN C N N yes TN D N N yes TN E P N miss FN confusion counts with P as positive count value TP 0 FN 1 FP 0 TN 4 metric fractions metric fraction accuracy 4/5 positive recall 0/1 = 0 accuracy can be high while the lone P is missed one pinned imbalanced table; one always-N baseline the lone P is missed; this is exact row counting does not claim: NOT fairness solved; NOT future performance NOT calibrated; NOT safe; NOT best metric; NOT accuracy guarantee NOT generalization; NOT probability truth
What this does not prove
It does not solve fairness, choose the best metric, prove safety, or prove future performance.
row-count arithmetic only \text{row-count arithmetic only} row-count arithmetic only Class Imbalance Exactly Exact row counts for one always-N baseline. label counts label count N 4 P 1 always-N baseline on one imbalanced table row true pred correct? cell A N N yes TN B N N yes TN C N N yes TN D N N yes TN E P N miss FN confusion counts with P as positive count value TP 0 FN 1 FP 0 TN 4 metric fractions metric fraction accuracy 4/5 positive recall 0/1 = 0 accuracy can be high while the lone P is missed one pinned imbalanced table; one always-N baseline the lone P is missed; this is exact row counting does not claim: NOT fairness solved; NOT future performance NOT calibrated; NOT safe; NOT best metric; NOT accuracy guarantee NOT generalization; NOT probability truth
The honest close
This is exact row counting only. It does not claim fairness solved, generalization, future performance, calibration, safety, best metric, accuracy guarantee, or probability truth.
exact row counts, narrow claim \text{exact row counts, narrow claim} exact row counts, narrow claim Class Imbalance Exactly Exact row counts for one always-N baseline. label counts label count N 4 P 1 always-N baseline on one imbalanced table row true pred correct? cell A N N yes TN B N N yes TN C N N yes TN D N N yes TN E P N miss FN confusion counts with P as positive count value TP 0 FN 1 FP 0 TN 4 metric fractions metric fraction accuracy 4/5 positive recall 0/1 = 0 accuracy can be high while the lone P is missed one pinned imbalanced table; one always-N baseline the lone P is missed; this is exact row counting does not claim: NOT fairness solved; NOT future performance NOT calibrated; NOT safe; NOT best metric; NOT accuracy guarantee NOT generalization; NOT probability truth