about the derivative calculation
ppen2013 opened this issue · comments
hello, i found that in the paper the loss function is log(e^-d(uv)/sum(e^-d(u,v'))).
it seems that the derivative over positive sample d() is -1, and the derivative over negative sample d() is e^d()/Z.
But in this C++ implementation , it seems not like that .
I was wondering whether i miss something.
looking forward your relpy
thanks
the variable u appears in both numerator and denominator, so the derivative is a little bit more complicated. I can help you if you write more concretely about your issues (with more concrete expressions) :)
Finally, I found the newest version of the paper from nips and the loss function interpretation was different from the one I saw before.
thank you all the same