the initial model for candidate selection was WRMF(weight regularized matrix factorization) from the name its pretty easy to assume what it’s supposed to be doing basically matrix factorization as used in regular collaborative filtering problems where the userxitem ie preferences/ratings or whatever floats your boat is factorized into user and item embeddings that can be used for subsequent predictions I think the regularization employed here is just for more efficiently learning (something i think is pretty relevant in this case since we’ve got like 20k….