add query-time pooling for flexMF

[samiravaez]

This PR resolves #934 by adding responsive recommendations to FlexMF.
FlexMF can now build a temporary user representation at scoring time by mean-pooling the embeddings of the items in RecQuery.query_items. This allows recommendations to be generated from query context even when there is no trained user embedding available.
Unknown query items are ignored, and if pooling cannot produce a usable vector, the model falls back to the trained user embedding when allowed. Explicit model scoring also preserves the global bias when using the pooled-user path

[mdekstrand]

We can just reuse the u_tensor, can't we? That will also enable making the if on line 172 simpler.

[mdekstrand]

If we reuse u_tensor above, then this can just be if u_tensor is None. I.e., the logic becomes "there are several ways to get u_tensor, if we've tried all of them and not gotten a u_tensor, then we can't score".

It is possible that this will also fix the test failure — I'm not sure if all of the if logic in this method is correct yet, partly because it seems more complicated than it needs to be.

[mdekstrand]

I'm not sure why we need a new function here - isn't the final result of pooling just a user embedding, like any other? That was the intention with the issue. (Whether it is a good user embedding is another question.)

[samiravaez]

what the user bias should be for the pooled embedding if I use the same scoring function?

[mdekstrand]

User bias is an important and subtle question here, and it might actually be different for explicit and implicit feedback models.

For implicit-feedback models, using 0 as the fresh user bias is probably fine. It might miscalibrate probabilities, but won't affect ranking.

For explicit-feedback models, I think we have 3 choices:

• use 0
• use saved user bias, even if we aren't using saved user embedding
• estimate user bias from rating values before computing the pooled embedding

[codecov]

Codecov Report

❌ Patch coverage is 97.67442% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 90.42%. Comparing base (34c81c5) to head (b038f74).
⚠️ Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
src/lenskit/flexmf/_model.py	91.66%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1201      +/-   ##
==========================================
+ Coverage   90.40%   90.42%   +0.01%     
==========================================
  Files         263      263              
  Lines       17516    17558      +42     
==========================================
+ Hits        15836    15877      +41     
- Misses       1680     1681       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.