macroF1

Macro F1, often written as macro F1 score, is an evaluation metric used to assess the performance of multi-class and multilabel classifiers. It measures the harmonic mean of precision and recall for each class, and then averages these per-class F1 scores to provide a single overall score that treats all classes equally.

To compute macro F1, for each class c, compute true positives, false positives, and false negatives using a one-vs-rest approach. Precision_c = TP_c / (TP_c + FP_c), Recall_c = TP_c / (TP_c + FN_c). If the denominator is zero, the corresponding value is defined as zero. The F1 for class c is F1_c = 2 * Precision_c * Recall_c / (Precision_c + Recall_c) when (Precision_c + Recall_c) > 0; otherwise F1_c = 0. Macro F1 is the average of F1_c over all classes: MacroF1 = (1/K) sum_c F1_c, where K is the number of classes.

Macro F1 is particularly useful when class distributions are imbalanced, as it gives equal weight to each class rather than to each example. It complements other metrics such as micro F1, which aggregates contributions across all classes before computing the F1 score, and weighted F1, which weights class F1 scores by class support.

Implementation and interpretation notes: In common libraries, macro F1 is produced by average='macro' in scikit-learn, or similar options in other frameworks. While easy to compute and interpretable, macro F1 can be influenced by rare classes with unstable estimates and does not capture the cost or importance of different misclassifications unless explicitly weighted.