Classical ML, validation, metrics, interpretation, debugging · insight-1182
Publish per-class recall with macro and weighted averages when a multiclass problem has rare classes
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- Data scientist, Data analyst
- Industries
- aviation, telecom, insurance
Summary: situation, action, reason, boundary
Situation
A multiclass model (disruption cause, claim type, complaint category) is summarized by accuracy or weighted F1, and one costly class holds a few percent of rows.
Action
Show per-class precision, recall and support in one table beside macro and weighted averages, let the cost of each class decide which summary is primary, and read the confusion-matrix rows of the rare classes.
Reason
Each averaging rule answers a different question: macro gives every class equal weight, weighted follows class presence, micro counts every sample-class pair.
Boundary
Multiclass classifiers with skewed class mix and unequal error costs.
Symptom or risk
Weighted averages follow class frequency, so the rare class can fail almost completely while the headline barely moves; macro averages over-emphasize a noisy low score on a tiny class.
Exceptions
Balanced classes, where the averages nearly coincide.
Trade-offs
A table is longer than one number; summaries for executives need a stated primary choice.
Pitfalls
Using micro-averaged F1 in single-label multiclass, where it equals accuracy.
Lessons and concepts
Claims and their sources
scikit-learn describes macro averaging as equal weight per class (it may over-emphasize low performance on an infrequent class), weighted averaging as weighting by class presence, and micro averaging as equal weight per sample-class pair; with all labels included micro precision, recall and F equal accuracy in multiclass.