This message was deleted.

I was reading this paper. https://dl.acm.org/doi/10.1145/3159652.3159687

I'm assuming that even though the recommendation system is 'offline', you're serving it to a user in some way that it can be tracked and desired outcomes of the recommendation are measured?

GrowthBook doesn't support experiments like that which use historical data instead of actually running against live users

ML model is being trained offline then it is being served to the live users

as long as you can see which user saw which model, it's possible to use GrowthBook to determine the difference