Diskussion: Wahl von Klassifikationsmetriken: Precision, Recall, F1, Schwellenwerte und Kalibrierung

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

Übersetzung nicht verfügbar; das Original wird angezeigt. Original

'A threshold-free score for comparing models' treats ROC AUC and average precision as interchangeable, and on the imbalanced problems the article uses as examples they are not. ROC AUC is built on the false-positive rate, whose denominator is the number of negatives; with 99 negatives per positive a model can produce many false positives per true positive and still show a false-positive rate near zero and an AUC above 0.9, while its precision at any useful recall is poor. Average precision has no such denominator and moves when the model's false positives change, which is what a fraud filter's operator feels. Saito and Rehmsmeier's 2015 comparison in PLOS ONE argued this point for imbalanced data, and it is why the class-imbalance article on this wiki names average precision specifically. The bullet should say: ROC AUC for balanced problems or when the ranking across all thresholds matters; average precision when positives are rare and the operating region is high precision. Reporting one 'threshold-free score' without saying which lets the better-looking number be picked.

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

Übersetzung nicht verfügbar; das Original wird angezeigt. Original

'The threshold is a deployment parameter, tuned on validation data, not part of the model' has direct support in scikit-learn since 1.5: `TunedThresholdClassifierCV` fits the underlying classifier and chooses the decision threshold by cross-validation against a scorer of the caller's choice (including a custom cost-based one), and `FixedThresholdClassifier` wraps a model with a threshold decided elsewhere, so the tuned cut-off is persisted with the pipeline instead of living in serving code. On the calibration side, 1.6 deprecated `cv="prefit"` in `CalibratedClassifierCV` in favour of wrapping the already-fitted model in `FrozenEstimator`, so examples that post-process an existing model with `cv="prefit"` warn on current versions.