ความแม่นยำ การเรียกคืน และ F1

To see past accuracy you split results into four buckets per class. From them come precision, recall, and F1, the metrics that matter.

Pick the class you care about as positive, like spam or fraud. Precision and recall both describe how well you handle that class.

Precision asks: of everything you flagged as positive, how many truly were? High precision means few false alarms.

Recall asks: of all the real positives, how many did you catch? High recall means few misses slip through.

Precision and recall usually trade off. Flag more items and recall rises but precision drops; flag fewer and the reverse happens.

Spam filters favor precision, since wrongly trashing real mail is costly. Cancer screens favor recall, since a missed case is far worse.

The F1 score blends precision and recall into one number using their harmonic mean, rewarding models that do both well.

F1 uses the harmonic mean so a model cannot fake it. If either precision or recall is near zero, F1 stays low too.

scikit-learn prints all three at once with a tidy classification report you can scan per class.

You can also pull a single metric directly when you only need F1 for the positive class.

No metric is best everywhere. Choose precision, recall, or F1 by the real-world cost of each mistake in your problem.

Match the metric to the question it answers.

Precision limits false alarms, recall limits misses, and F1 balances both. Let the cost of each error pick your metric. 📊