Today I held a short laboratory which tackled different metrics used in evaluating classifiers. One of the tasks required that, given the performances of 2 classifiers as confusion matrices, the students will calculate the accuracy of the 2 models. One model was a binary classifier and the other was a multiclass classifier.
Many students resorted to googling for an accuracy formula which returned the following function:
Then, they calculated a ‘per-class’ accuracy (for class , they had ) and macro-averaged the results like below:
To their surprise, the resulted accuracy for the multiclass classifier was erroneous and highly different (when compared to
accuracy_score() from sklearn). However, the accuracy of the binary classifier was correct.
As there wasn’t much time available, I told them to use the following accuracy formula to match the results of sklearn and I’ll send an explaination later:
Some of you might recognize this as micro-averaged precision.
The purpose of this article is to serve as a list of DO’s and DONT’s so we can avoid such mistakes in the future.
What was wrong?
Basically, you’re prone to get invalid results if you average accuracy values in an attempt to obtain the global accuracy. But… even if you directly calculate the global accuracy using the above formula, you’d get skewed values.
Take a look at the following classifier, described using a confusion matrix:
|\||Class #0||Class #1||Class #2|
You’ll notice that thus the classifier is doing a really bad job.
If we follow the students’ approach and calculate the ‘per-class’ accuracy (let’s say Class #0), we have:
This already looks suspicious. You’ll get the same results for the other 2 classes, so… on average, . This is definitely wrong.
If you directly compute global accuracy using the same formula (summing all , , …), you get the same result because of the symmetry. This happens mainly because of the in the numerator which grows faster than any other term. In other words, as the number of classes grows, this error grows aswell; a similar model, but with 4 classess, gets a 0.5 accuracy.
Using the second formula, the global accuracy becomes:
Which yields, indeed, a better result. Moreover, it generates the same results as
accuracy_score() from sklearn, given more diverse confusion matrices.
If you compute ‘per class’ accuracies using the second formula and average the values, you’re basically getting a macro-averaged precision. Point is, that’s not accuracy - so don’t do that.
I’d recommend avoiding:
- the idea of calculating a global accuracy by averaging ‘per-class’ accuracies
- the red formula, which includes , since the other one returns correct values for any number of classes
Overall, you can compute precision, recall, F1 in a ‘per-class’ manner. But I’m not so sure it also works with the accuracy.