Setup:
I need to build a ranking model to rank hotels (let’s consider a search system without personalization), so my final unit of observation in the app must be a hotel. But during training&testing my model I’ve got a client query as a unit of observation (so my data looks like this: query_id + hotel + features + target).

Question:
What’s the best way to go from the initial unit of observation to the final one? Is it a good idea to aggregate results grouping by hotels after making predictions for each query and calculating mean/median score? Any ideas?

Thank you in advance~~

Now the only idea I’ve got is to calculate mean/median predicted scores in the Catboost model grouping them by hotels. But it seems to be not the best one…