Predict recommendations without retraining model for new users

Hello!

I'm looking for a way for effective item recommendations for new users. I've read all created issues about LightFM (the most relevant existing question is https://github.com/lyst/lightfm/issues/347), but have not found full answer on my question.

My situation is:
I have historical data, where users interact with subset of items from finite set. Users have feature representation (gender, age, etc), items have features (tags, types, etc). So, in my opinion LightFM hybrid model will be very suitable for giving user recommendations about items.
In production mode item set will be the same. BUT all users will be new. I will know user features. And I need in online mode take new user-item interaction and give new recommendations. Is it possible to take into account new interractions in realtime? How? Is it possible to get this recommenations without changing model (i want to "freeze" model)?

Thank you any help!

Start by reviewing LightFM's existing hybrid-model APIs and the discussion in issue #347. Investigate whether known user features and new interactions can support realtime recommendations while the trained model remains frozen. Done means establishing the supported workflow or the missing capability and documenting the findings clearly.

Written by the indexing model from the issue text.