Learning curve using micro F-score and macro F-score

Question

I plotted the learning curves using micro and macro F-scores for a Multinomial Naive Bayes classifier.

The first plot is made using micro F-score, and the second using macro F-score. I find it quite difficult to interpret both of them. The explanations I found for learning curves use errors.

I think in the first plot there is variability up to around 300 instances, and then it starts to converge until the point when the lines are parallel. Therefore, even adding more data would not help. But what about bias?

Also can you please explain to me the difference between the plots in terms of micro and macro F-scores?

Answer 1

Micro calculates F score globally by counting the total true positives, false negatives and false positives.

Macro calculates F score for each label and find their unweighted mean. Macro F score does not take label imbalance into account.

Given there is a difference in your performance between the metrics, your data is imbalanced in the base-rate for the class labels.

Adding more data to the minority class(es) will help.