Bug in RankALS, Q-Step about librec HOT 3 CLOSED

Thanks for your feedback.

After checking the code and the original paper, the code below

DenseVector pu = P.row(u);
double rui = trainMatrix.get(u, i);
DenseMatrix pp = pu.outer(pu);
sum_cpp = sum_cpp.add(pp);

says sum_cpp is the sum of p_u p_u^t, i.e., the \bar{A} in the paper. Since \bar{A} is not dependent on r_ui, I don't know why you suggest to put it in the if condition.

What do you think?

from librec.

Sorry, didn't see that you responded!

It is written in the paper that \bar{A} = P_{U_i}.dot(P_{U_i}.T), the U_i means only all users that gave a rating to item i. Therefore, only if r_ui is larger than 0, meaning that we have a rating, we should add them to \bar{A}. Otherwise, you could precompute \bar{A} and don't need to do it in every item step. I added below the important section of the algorithm as an image. Hope my explanation helps.

from librec.

@466152112 @wubin7019088 what do you think?

from librec.