How to Calculated Highest-Recommended item(s) across groups of users/rows?

I have a set of LightFM recommendations I have trained via WARP loss and just implicit feedback (user-item transaction history; no separate user-features or item-features yet).

I have groups of users/rows that are relevant for my business case, and I want to calculate the "most-recommended (or top 3) items across [group of users]." However, I've seen warnings in other comments that recommendation scores seem not to be comparable across user/rows (e.g. [maciejkula](https://github.com/maciejkula) wrote "User-item scores out of the LightFM model do not have a guaranteed range or scale (as that is irrelevant for ranking items for any particular user)" in 2018 in [this issue](https://github.com/lyst/lightfm/issues/242)

Is there a recommended way to calculate e.g. top-1,2,3 recommended items across groups of users, since we shouldn't average user-item recommendation scores across user/rows? A couple possible ways forward below, but wanted to ask the experts in case I'm off-base somehow:
- Normalizing user-item recommendation scores within rows (so each user/row's item scores vary -1 to 1), and then averaging those normalized values across rows (maciejkula recommended something like this in the issue above, albeit in response to a different usecase/question)
- Just counting the number of users an item is top-ranked for. Top-ranked item across users is item that gets most highest-recommendation-score. (Though this would not provide credit/weighting for items which are consistently e.g. 2nd-highest-recommended instead of 20th-highest-recommended item for many users).

Any advice/guidance or corrections to anything I may have misunderstood would be much appreciated. Thanks for maintaining this great library!

No source file, test, or entry point is named. Start by reviewing LightFM’s user-item score semantics and the aggregation approaches described in the issue; done would require a maintainer-backed recommendation for a group-level ranking method, with scope and validation criteria recorded.

Written by the indexing model from the issue text.