I'm pleased to announce the release of LensKit 0.12. This version brings mostly some internal and distribution management improvements, but we're no longer carrying around our own sparse matrix class - that is spun out into the brand-new CSR library. We have also moved to Flit for distribution, with a 'flit-conda' script in build-tools/ to create a Conda development environment from the Flit specifications in pyproject.toml.
One important change for any of you who may have used our TF BPR: it now defaults to weighting negative samples by popularity instead of sampling them uniformly at random. Thanks to Karl Higley and Ben Frederickson for explaining the value of this to me at RecSys 2020. Note that we don't have a lot of experience with this implementation yet, so there might still be some kinks to work out.
You can install LensKit 0.12 from conda-forge or from PyPI.
--
I may send mail outside of working hours; I do not expect you to. He/him.