Add IPS-weighted RBP (#1155)

[vvineis]

Adds IPS-weighted RBP (#1155), following Yang et al., "Unbiased Offline Recommender Evaluation for Missing-Not-At-Random Implicit Feedback" (RecSys 2018).

The metric is IPSRBP. It uses the paper's self-normalized estimator (SNIPS, their eqn. 9) by default; self_normalized=False drops the denominator and gives the plain IPS numerator. Propensities come from a PropensityModel passed to the constructor: PopularityPropensity implements their eqn. 15, UniformPropensity is the missing-at-random baseline, and FieldPropensity reads them off a field if you estimated them yourself. There's also estimate_power_law_gamma, which fits the exponent from a set of recommendations the way they do in section 4.3.

I put the propensities in a separate object rather than as arguments on the metric, partly to follow MeanPopRank and ListGini taking a Dataset, and partly because it mirrors RankWeight. IPS-corrected Recall or DCG should then just be a matter of reusing it.

A few choices:

• gamma is the paper's exponent, so the (gamma+1)/2 of eqn. 15 happens inside the propensity model and values from their Table 2 can be used as-is. It has no default, since it depends on the dataset and on whatever produced the log.
• The paper only pins propensities up to a constant. I normalize by the most popular item, so P lands in (0, 1] and weights are at least 1. It makes no difference under SNIPS, but the unnormalized form needs a fixed scale.

Tests

Two identities tie the metric to the existing one, both checked with Hypothesis over random lists, cutoffs and patience values:

• uniform propensities with self_normalized=False is exactly RBP
• uniform propensities with SNIPS is exactly the paper's average-over-all evaluator (eqn. 6), which for RBP works out to RBP / |S*_u|

I also checked both per user on ml-latest-small and they hold to machine precision.

There's also a small simulated MNAR log where the true reward of eqn. 1 is known: the uncorrected evaluator overstates a popularity-chasing recommender and the correction pulls it back, while on a missing-at-random log the correction leaves the score alone.

One thing I didn't do

The docstrings cite yangUnbiasedOfflineRecommender2018, which isn't in docs/lenskit.bib yet.

@inproceedings{yang2018unbiased,
title={Unbiased offline recommender evaluation for missing-not-at-random implicit feedback},
author={Yang, Longqi and Cui, Yin and Xuan, Yuan and Wang, Chenyang and Belongie, Serge and Estrin, Deborah},
booktitle={Proceedings of the 12th ACM conference on recommender systems},
pages={279--287},
year={2018}
}

[codecov]

Codecov Report

❌ Patch coverage is 95.77465% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 89.63%. Comparing base (024fc7b) to head (de8fb82).

Files with missing lines	Patch %	Lines
src/lenskit/metrics/ranking/_propensity.py	94.68%	5 Missing ⚠️
src/lenskit/metrics/ranking/_ips.py	97.82%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1182      +/-   ##
==========================================
+ Coverage   89.58%   89.63%   +0.05%     
==========================================
  Files         262      264       +2     
  Lines       17405    17547     +142     
==========================================
+ Hits        15593    15729     +136     
- Misses       1812     1818       +6     

☔ 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.

[mdekstrand]

Will provide a more detailed review tomorrow or Monday, but I just cut the 2026.3 release, so this will be targeted for 2026.4.0.