SocialMF update rule for P about qrec HOT 4 CLOSED

In this paper, the batch gradient descent is used to look for the local minimum solution. However, in our implementation, we used the stochastic gradient descent. So the last term in this formula is omitted. Besides, we use the max-min normalization instead of g(x) to save time because calculating g' is time-consuming. In fact, the way of normalization is small-beer.

from qrec.

I am sorry to ask, I am not familiar with the method you described, could you please reference a lecture note for that particular method?

I think that with any method we use we can't omit that missing part, because if we want to calculate gradient of L against U_i , U_i appears in other terms(for example U_i is neighbor of U_k).

from qrec.

When using the stochastic gradient descent, we choose a tuple from the whole dataset randomly. Each time we only consider one tuple. The optimization function and gradient will be changed into the equations below.

Hope it can help you.

from qrec.

Ok, I got it, I close this issue then.

from qrec.