Micro-averaging pools predictions across all classes or labels before calculating one metric. Because frequent labels contribute more observations, the result emphasizes overall instance-level performance and is useful for imbalanced multilabel activity data. section 5.4.2
Use micro-averaging when evaluation should reflect the labels occurring most often:
- Pool true positives, false positives, and other outcome counts across classes, then calculate the metric once. section 5.4.2
- Prefer it when common activity labels matter most or when overall multilabel performance matters more than equal performance across labels. section 5.4.2
- Choose macro-averaging instead when rare labels deserve the same influence as frequent labels. section 5.4.2
Pooling three class results
Suppose three classes produce 90, 30, and 40 true positives, with 10, 20, and 10 false positives. Micro precision pools the counts: 160 / 200 = 0.80. Macro precision averages the class precisions, (0.90 + 0.60 + 0.80) / 3 ≈ 0.77. The higher micro value reflects the influence of the large, high-precision class. section 5.4.2
