mikro in perfomancevector

nguyenxuanhau

hi!
how was mikro calculated in performancevector?
Thanks

IngoRM

Hi,

the micro average is simply the average of the performance of all examples. The macro average is only applicable in cases where the examples are divided in batches like it is done by cross validation. In this case, the macro average is the average of the performance of all batches where each batch performance is calculated as average of the examples' performances.

Here is an example: assume that you have 200 examples which are divided by cross validation into 10 folds with 20 examples in each fold. The micro average is now simply the average of all the performance of all 200 examples. The macro average is the average of the average performance of each of the 10 folds. Each of those fold- or batch-averages is calculated basically as a micro average of the 20 containing examples.

Hope that helps,
Ingo

FelipeMorais

Ingo, I saw your answer to this question, but I could not generalize it. In my case, I am using 10-fold cross-validation on batch and the performance metric I am using is Kappa. My models detect the presence or absence of a certain state from students. However, for a given state I have gotten kappa: 0.094 +/- 0.266 (micro average: 0.421). The difference is too big. Thus, I would like to know which one should I report 0.094 or 0.421?  Thank you for your help.

IngoRM

Hi,

The macro average.  The whole point about the 10-fold cross validation is to get to the macro average.  So if you report / compare performances created by cross validations, you should always use the macro average.

Hope this helps,

Ingo

FelipeMorais

Alright. Got it. Thanks for your help and the fast reply.
Best regards.
Felipe.