Upselling recommendation with NDR for implicit feedback (millions users, few items)

I am currently developing an upselling recommendation system.

Scenario:

• Millions of users
• Few dozen items
• Most users do not interact with any items
• No ratings, only known whether the user has interacted with the item (implicit feedback)

Approach:

• Implementing an NDR (Two-tower) architecture using user and item features
• Currently treating it as a retrieval problem, inputting only user-item interactions (positive feedback) to the model

I was wondering if this is the right approach or if it might make sense to transform it into a ranking problem (1 if the user interacts and 0 otherwise), with the caveat that the size of the database could explode because if I understand correctly, negative examples would have to be created manually.
What is the best way to proceed?

The issue names no file, test, or entry point. Start by reviewing the TensorFlow Recommenders retrieval and ranking examples relevant to implicit feedback and two-tower models, then clarify whether the expected outcome is implementation or guidance. Define a concrete success criterion and get maintainer direction before making changes.