Reference · Glossary

Statistics for ML

Last updated

Evaluating classifiers, comparing A/B model tests, checking class imbalance, reporting precision/recall, and interpreting benchmark tables.

#When to use

Evaluating classifiers, comparing A/B model tests, checking class imbalance, reporting precision/recall, and interpreting benchmark tables.

#When not to

As a substitute for domain review — a statistically significant metric can still be useless if it ignores business cost.

#Example

1000 emails, 50 spam (5%)
Model always predicts "not spam" → 95% accuracy but catches 0 spam
→ need precision, recall, and confusion matrix — not accuracy alone

Pair summary stats with **splits**, **confidence**, and **error analysis**.