희소 사건에 맞는 평가지표 선택하기

You already distrust accuracy on imbalance. Now you need metrics that actually reward catching the rare event.

Precision is the share of flagged positives that are truly positive. High precision means few false alarms among your alerts.

Recall is the share of real positives you actually caught. For rare, costly events, missing them is the worst outcome.

The F1 score blends precision and recall into one number. It rises only when both are decent, punishing a lopsided model.

A precision-recall curve traces the trade-off across every threshold. It speaks directly to the rare class you care about.

On heavy imbalance, ROC curves can look rosy because true negatives are plentiful. The PR curve stays honest about the rare class.

Average precision summarizes the PR curve in a single score. A higher value means strong precision held across many recall levels.

One scikit-learn call prints precision, recall, and F1 for every class at once, so the minority is never hidden.

In that report, fix your eyes on the rare class row. Its recall and precision tell you whether the model truly works. 🎯

If missed cases are dangerous, optimize recall. If false alarms are expensive, favor precision. Let the real-world cost decide.

Never lean on a single number. Show precision, recall, F1, and a PR curve together so nobody is fooled by inflated accuracy.

Which metric best suits rare, important events?

For rare events, judge with precision, recall, F1, and PR curves, not accuracy. Match the metric to which mistake costs you most. 🎯