In imbalanced NLP classification tasks, why use macro or micro averages in confusion-matrix-based metrics?

• A. Macro averages give equal weight to all classes; micro averages aggregate counts across all samples, which can be dominated by the majority class.
• B. Macro averages ignore minority classes.
• C. Micro averages are always better for imbalanced data.
• D. Macro averages only apply to regression.

Answer: (A)

When evaluating imbalanced NLP classification, you want metrics that reflect performance across all classes, not just the majority one. Macro averaging does this by computing the metric (precision, recall, F1) separately for each class and then averaging those values. This gives equal weight to every class, so the model’s ability to handle minority classes is directly visible in the score. If a model struggles with rare categories, macro averages will reveal that even if overall accuracy looks decent.

Micro averaging, on the other hand, pools all correct and incorrect predictions across all classes before computing the metric. Because there are many more instances from the dominant class, micro averages tend to be driven by that class and can mask poor performance on minority classes.

So macro averages are best when you want balanced attention to every class, while micro averages reflect overall performance across all predictions. In imbalanced NLP tasks, it’s common to report both to get a complete view.