Closed CicoZhang closed 6 years ago
These two modes have the different objective functions. “single” mode uses the negative log likelihood of sigmoid activation as an objective, and “pairwise” mode uses max-margin loss.
In my opinion, there isn’t the significant difference in performance between single and pairwise mode, but it may depend on the model or dataset you use.
Great, thank you for the explanation!
I am wondering if you could give a simple explanation of the mode "single" and "pairwise". Is there any preference to stick to one mode?
Thank you in advance.