Robs opinion on which metric to use for classification tasks
I commonly come accross evaluations where in my opinion the wrong metrics are used, even in some prominent benchmarks. This seems to be mainly due to the misundertanding that for datasets/tasks with unbalanced class-distribution, macro-F1 is always the correct metric. I disagree with this heuristic. What is more important is to ask yourself: do I care most about all classes being found correctly, or do I care more about the number of correct instances. Furthermore, accuracy has the benefit of being more interpretable, having a value of 0.66 just means we got 66% of all instances correct, but for F1 this is less obvious, it could be recall = .6 and precision is .6, but they could also be .5 and 1.0. Of course, we do need to include the majority baseline for interpretation of accuracy. Another strange combination I see too often is a binary task with macro-F1; it should be noted here that an error for class A is automatically also an error in class B (an FP for A becomes a FN for B), which is probably not desired (note that SKLearn gives a warning for this).
After discussing with some of my collegues (mostly Elisa and Christian), and mostly agreeing to disagree, I have came up with a decision tree (shown below). It should be noted that this is subjective (if I didn’t make this clear enough yet).
Comments