Yes, I also don’t find it clear how the rating is used.
It’s clear that the rating data is passed as the input here
compute_loss code in the “combined model” section doesn’t use the
It calculates some loss based on the features… I think the clue is the
self.task = tfrs.tasks.Retrieval.
Retrieval and rating are often two separate steps. IIUC Retreival is “what should I even borther running the rating model on?”. So I think this model is very similar to any collaborative filtering model, except that instead of trying to predict what the rating is, it’s trying to predict “would this user have rated this movie in the first place?”.
I’ll ping the owner and see if they have any feedback.