I haven’t used the CollabFilterDataset API for generating test predictions, but I have used similar model in different competition with some success. (I would love to know if someone has used this API to make predictions, especially the test_df parameter; or maybe the predict method.)
AFAIK, for those variable combinations where we have data, we can just dot product corresponding embedding vectors to get the rating.
In case of predictions for variables for which we don’t have any data - for example in case of new users - I think one way would be to start off with median/ mean movie ratings for combination of this user and respective movie. If it’s both new variables, we can start off with some other sensible value to start off with.