How do people actually generate unseen new inference data?

Describe the question(问题描述)

Let's say I implemented DeepFM here. I see how we train and test. But after we test, now we want to actually make recommendations with unseen new data. But how do we then create the unseen new data? What is people normally do? Because there are tens of millions users and millions movies, I don't think people generate unseen new data point for each user to go on for loop of every movie. Would you tip me off how people generally do in production?

No file, test, or entry point is named in the issue. Start by locating the repository's DeepFM implementation and any inference or usage documentation, then confirm with maintainers whether documenting the production workflow is the intended scope and what completed guidance should cover.