Cheat sheet

Model Evaluation Metrics

Confusion-matrix metrics, ROC and PR curves, calibration, thresholds, multiclass averaging, regression, ranking, and reporting.

	Predicted positive	Predicted negative
Actual positive	TP	FN (miss)
Actual negative	FP (false alarm)	TN

Average	How	Use when
Micro	Pool all TP, FP, FN	Overall performance; equals accuracy for single-label
Macro	Mean of per-class scores	Every class matters equally
Weighted	Per-class scores weighted by support	Reflect class frequency

Situation	Primary metric
Balanced classes, equal costs	Accuracy, F1
Rare positives	PR-AUC, recall at precision ≥ X
Ranking quality only	ROC-AUC
Decisions need probabilities	Log loss, Brier, calibration
Regression with outliers	MAE or Huber
Search and recommendations	NDCG@k, MRR, recall@k