Two reports can use the same predictions yet show different averages. The aggregation method determines which categories have more influence.
Scikit-learn distinguishes macro averaging of class-level metrics from micro averaging across pooled decisions. Scikit-learn — model evaluation.
Preserve the smaller categories
For a character-detail task, an uncommon accessory category may matter even when it contributes few examples. Show its own results instead of relying only on an aggregate dominated by common categories.
Choose before comparing
State which average supports the practical question and retain category counts. Changing the averaging method after seeing the result can create a misleading improvement story. The preferred average should follow the decision's priorities, with those priorities made explicit.
Source reviewed 25 October 2026. Interpretation and original diagrams: PLUSHIP / Qunhe. Examples are illustrative.
Related reading
Semantic and Instance Segmentation Organise a Scene Differently
Hyperparameter Search Needs a Separate Performance Estimate
Object Detection Can Locate a Detail Without Measuring Its Quality