For highly imbalanced fraud detection, which metric set is MORE app…

Question

For highly imbalanced fraud detection, which metric set is MORE appropriate than plain accuracy?

Accepted Answer

Precision, recall, F1, PR-AUC (and careful thresholding) — Imbalanced fraud: precision/recall/PR-AUC beat vanilla accuracy every time.

Answer

Accuracy alone on 99.9% negative class

Answer

Number of Spark partitions only

Answer

Only training loss with no validation metrics

For highly imbalanced fraud detection, which metric set is MORE appropriate than plain accuracy?