What metrics of multiclass classification do you know? Their pros and cons (macro, micro, weighted).

In multi-class classification tasks, the following aggregation metrics are often used:

• Macro-average — calculates the metric (e.g., precision, recall, F1) for each class separately, then averages them without considering the number of examples in each class.

  • Pros: gives equal weight to all classes, useful for imbalanced data to not ignore rare classes.
  • Cons: can be too sensitive to classes with few examples.

• Micro-average — combines all predictions and true labels across all classes and computes the metric on this combined set.

  • Pros: reflects the overall performance of the model, considering the number of examples in each class.
  • Cons: dominates classes with many examples, rare classes may be "dampened".

• Weighted-average — computes the metric for each class and averages them, weighted by the number of examples in each class.

  • Pros: balances between macro and micro, considering each class's contribution proportionally to its size.
  • Cons: may still underestimate rare classes if they are very small.

The choice of metric depends on the task: if equal evaluation for all classes is important — macro, if overall balance — micro, if a compromise is needed — weighted.

---

Example in sklearn:

from sklearn.metrics import precision_recall_fscore_support

# y_true and y_pred are true and predicted labels
precision, recall, f1, _ = precision_recall_fscore_support(y_true, y_pred, average='macro')