ATLASResearch
methods
Quantitative/ Performance analysis

ROC-AUC and precision-recall curves

Examine ranking across decision thresholds

ROC curves show true-positive against false-positive rates over score thresholds; ROC-AUC summarises ranking discrimination. Precision-recall curves instead expose positive prediction quality and coverage and depend on prevalence. Average precision and trapezoidal PR-AUC use different summarisation rules and should not be conflated. Curves do not assess whether predicted probabilities are calibrated.

WHEN IT FITS

Use curves when comparing scoring models across plausible thresholds, especially when deployment thresholds are not fixed, while considering the class balance and operating region relevant to the intended use.

Strengths

  • Shows threshold trade-offs that one cut-off hides
  • PR curves are informative when positive predictions must be trustworthy

Limitations

  • A whole-curve average can conceal poor performance at the required operating point
  • ROC-AUC does not express probability calibration or deployment utility

Know the boundary

High ROC-AUC does not guarantee a high precision at the threshold you will deploy.

USED ACROSS
Computer scienceBanking & financePsychologyEducation