I may be misreading or oversimplifying, but I think what he is talking about is fold-in. Given the item-feature matrix, there is a simple analytical solution for the new user-feature matrix row which you compute directly. Or else, I don't know what else is intended... he's not talking about updating anything item factors or any other user features after all.
On Fri, Oct 12, 2012 at 6:19 PM, Ahmet Ylmaz <[email protected]> wrote: > Hi Sebastian, > What do you think of the technique of retraining the model *only* on the new > user, in order to learn the factors associated with him. > (Yehuda Koren mentions it here: > http://www.netflixprize.com/community/viewtopic.php?id=815) > > Is it worse than folding-in?
