Accuracy 和 Recall 值相同

Question

I have trained CNN models using Pytorch with Python programming language, I try to obtain metrics from the test data set by using sklearn.metrics as shown below. But I got same result for accuracy and recall. Is there any best practices to show metrics? Does this result true?

test_accuracy_score = accuracy_score(output_list, prediction_list)
test_precision_score = precision_score(output_list, prediction_list, average=‘weighted’)
test_f1_score = f1_score(output_list, prediction_list, average=‘weighted’)
test_recall_score = recall_score(output_list, prediction_list, average=‘weighted’)

Answer 1

If you want a detailed result just import classification report and print :

print(classification_report(output_list, prediction_list)) which also shows the support for each class (number of instances).

Your weighted recall returns the recall for each class adjusted for the number of elements on each class. So we have accuracy which is defined as:

acc = (TP + TN) / (TP + TN + FP + FN) -- T for true P for positive etc...

and recall defined for classA and classB (let's assume binary classification)

recall1 = TP / (TP + FN) for class1

recall2 = TP / (TP + FN) for class2 which translates to TN / (TN + FP) (if you imagine that class1 is positive and class2 is negative)

weighted recall returns the number

w_recall = ElementsOfClass1overAll * recall1 + ElementsOfClass2overAll * recall2 = ElementsOfClass1overAll * (TP / (TP + FN)) + ElementsOfClass2overAll * (TN / (TN + FP))

So, we can say that if ElementsOfClass1overAll and ElementsOfClass2overAll are equal then w_recall is exactly equal to your accuracy (meaning if half the elements belong to class A and half to classB).