Use sigmod or tempered_sigmoid for prediction?

[jamesguoxin]

Dear authors,

I'm encountering an issue regarding to the prediction phase. Assume we use bi_tempered_binary_logistic_loss as our loss function, should we still use sigmod as function used for probability calculation, or we should use tempered_sigmoid? Thanks for clarification!

[eamid]

You should use the tempered_sigmoid function to calculate the class probabilities. However, to predict the class, you don't need to form the probabilities (i.e. pick the class with the larger logit value).

[jamesguoxin]

Hello eamid,

Thank you for your reply. In our case, we would like to extract the probability to calculate AUC for the prediction. So for the t parameter in tempered_sigmoid, we should keep it consistent with bi_tempered_binary_logistic_loss t1?

[eamid]

It should be consistent with t2. The other temperature t1 controls the boundedness of the loss

[jamesguoxin]

I see, thank you eamid!