I’ve read that once you train your collaborative filtering model, depending on how much users interact with it, it can start to decay really fast. Also updating it for every new interaction can be really expensive.
Imagine you are putting into production models for multiple e-commerces. What are some good techniques to know when your interaction dataset has drifted enough so that it will be good to trigger a retraining process for one of the models? So you retrain it only when it is really needed to (maybe months later if the data does not drift a lot, or every week if customers behavior drifts regularly)