[Discussion] Handle user ID with continuously training

I want to open a discussion here as I feel the tutorial is good for running a small dataset, a one-pass model. However, in the deployment, there are more processes needed.

One thing I found might be useful is using the hash trick, it basically maps customer id to a fixed-length category bin (let's say 500K). The benefit is now we will have a fixed length parameters set. Then, I am thinking of continuous training, we could do a warm start with the previous model and train with new samples.

Want to see what this community thinks about this solution and any other ideas :)

Start by reviewing the tutorial referenced in the discussion and the linked tf.keras.layers.Hashing API. The issue does not name files, tests, or a concrete implementation target, so first clarify the expected continuous-training workflow, warm-start behavior, and acceptance criteria. Work would be complete only once the community agrees on a specific, testable change.