best empirical macro/micro F1 score? In the following presentation it's said that "0.5 to 0.55 (micro) F1 score is best for multilabel classification problems"
I tried to investigate this statement but couldn't find the source.
Does anyone know if there is more to it?
 A: On an extremely easy data set, anything but a F1 of 1.0 is a failure. ;-)
On a data set, where $l$ labels are assigned at random with the same probability, precision and recall of much more than $1/l$ are either due to chance or due to cheating.
Real data will of course be somewhere inbetween.
Now here is the catch: if you have a balanced, binary classification problem, and answer at random. You have a chance of 50% to get the values right, so on a balanced, binary classiciation problem, F1 of 0.5 is the expected performance of a random classifier. So any result around 0.5 is useless. This does, however, not generalize to balanced multiclass problems!
So maybe that was the statement where that number came from: 0.5 on a binary problem is random; but 0.5 on a 100-class problem can be pretty good (if it is a balanced, difficult data set without issues such as duplicates etc.). The statement probably never was 0.5 is "best".
So since a random result is expected to achieve $1/l$ as F1-measure, you could define an adjusted F1 measure that is less sensitive to the class size:
$$adjF1_l:=\frac{F1 - 1/l}{1 - 1/l}$$
But this equation assumes that all classes are equally frequent, so it is not a general purpose thing.
A: I think you're missing the context here: He says that results reported in the literature on this task (multi-label document categorization) are typically around 0.5-0.55, but the Kaggle leaderboard had some people reporting results that were much higher (~0.8).  He then discusses why (the test set contained copies of training examples) and how he built a system that used that information, plus other data, which performed really well. 
He doesn't mention where he got that "typical" range from, but it's not totally unreasonable. This paper by T .Rubin et al. reports pretty similar F1 values in Table 8, but the numbers in this comparison of two systems are a bit higher. 
There's nothing magical about 0.5. Depending on the task, a number much less than 0.5 might be a cause for celebration or even something like 0.95 might be disappointing. 
