metric.md

ROC (Receiver operating characteristic) & AUC (Area Under Curve)

• TPR (y-axis) vs FPR

# Alis for recall or sensitivity
TPR = TP/P = TP / (TP + FN)

# Alis for fall-out
FPR = FP/N = FP / (FP + TN)

Precision-Recall & AP (average_precision_score)

• AP is a summary of Precision-Recall curve
• precision (y-axis) vs recall
• Useful when classes are unbalanced.

Precision = TP / (TP + FP)

Recall = TP / (TP + FN)

# area under Precision-Recall curve
AP = sum_n (R_n - R_{n-1}) P_n

multi-label average

micro: ignore classes, directly calculate metrics.
macro: calculate metric per-classes, then average the metrics.
mAP: mean Average Precision, claculate AP for each class, then average over classes.
mAUC: claculate AUC for each class, then average over classes.