Mathematically, the micro-averages measures in mulan, should be equal to the corresponding example-based ones. However mulan usually outputs slightly different values. Why does this occur? Is it in the way divisions by zero are handled in precision/recal calculation?
Nobody/Anonymous
2008-12-17
The divisions by zero seem to be handled properly in precision/recall calculation.
tsoumakas
2008-12-17
My conclusion is that they are mathematically equal provided no division by zero occurs in the example-based metrics (where the probability of division by zero is higher) so the results are fine. There is no reason to delete one or the other, because they are also conceptually different. Biologists for example, understand the example-based, while micro-averaging is counter-intuitive.
Anonymous