scikit-learn classification report's f1 accuracy?

Question

When I run scikit-learn classification_report() on my 2-class y and yhat, I get the following:

              precision    recall  f1-score   support
       0       0.84      0.97      0.90    160319
       1       0.67      0.27      0.38     41010

accuracy                           0.82    201329   <--- WHAT?

macro avg       0.75      0.62      0.64    201329
weighted avg       0.80      0.82      0.79    201329

I understand F1, macro-averaged F1, etc. But what is the accuracy value of 0.82 under the f1-score column?

I already read other questions about this function:

How to interpret classification report of scikit-learn?

Sklearn classification report is not printing the micro avg score for multi class classification model

macro average and weighted average meaning in classification_report

Answer 1

I give you that this is a weird way of displaying the data, but the accuracy is the only field that don't fit the schema. For example:

              precision    recall  f1-score   support
       0       0.84      0.97      0.90    160319
       1       0.67      0.27      0.38     41010

As explained in How to interpret classification report of scikit-learn?, the precision, recall, f1-score and support are simply those metrics for both classes of your binary classification problem.

The second part of the table:

    accuracy                           0.82    201329   <--- WHAT?
   macro avg       0.75      0.62      0.64    201329
weighted avg       0.80      0.82      0.79    201329

The accuracy is the overall accuracy of the model (note that accuracy is not a measure that is relative to a certain class, but a performance across all classes). The macro average for the precision and recall score is just the harmonic mean of the two classes. ie:

recall macro avg = (recall_class_1 + recall_class_0) / 2

The weighted average is just the average metric of the two classes weighted by the support/size-of-sample. ie:

recall weighted avg = (support_class_0 * recall_class_0 + support_class_1 * recall_class_1) / (support_class_0 + support_class_1)

This is a pretty long-winded way of saying the accuracy is just the overall accuracy. It has nothing to do with the column f1-score that it happens to be under.