2026 Releases#

2026 builds on the foundation of 2025 to improve the ergonomics of data access, querying, and metrics, and clean up some APIs that seemed good at the time, while making it even easier to use LensKit for recommendation scenarios besides ID-based personalized recommendation.

There are no new major paradigm shifts, though — pipelines, datasets, and components work as they do in the 2025 series, but with more features, some rough corners polished off the interfaces, and hopefully fewer bugs.

2026.0.0#

Breaking Changes#

  • The Trainable protocol now has a separate method is_trained() to query if a component has been trained, and responsibility for skipping retraining of already-trained components when retrain is False has been moved from individual components to lenskit.pipeline.Pipeline.train() (🐞 1042, ⛙ 1044).

    The pipeline builder will issue a warning if a component implements train() but not is_trained().

  • Pipeline configurations serialized with previous versions cannot be re-loaded in LensKit 2026, due to moves of module paths. Import path canonicalization (🐞 948) reduces the risk of such breakage in future releases. Handwritten configurations will often still work.

  • Many submodules (such as most modules under lenskit.pipeline) have been renamed to be private modules. Code importing from their original locations will need to be updated to import from the higher-level module (🐞 947).

  • LensKit now requires Python 3.12 or newer, along with NumPy 2.x, Pandas 2.3 or newer, and SciPy 1.13 or newer (see Dependency Versioning, ⛙ 954).

  • We no longer publish 32-bit binary wheels.

  • Removed DecomposedMetric, as the Metric interface is now decomposed. All metrics based on listwise measurements or intermediate results should directly extend from Metric or ListMetric. (⛙ 983)

  • GlobalMetric no longer inherits from Metric, and may be removed in a future release.

  • Stopped providing wheels for macOS on Intel. Users who still need to run LensKit on Intel-based Macs should use the Conda packages (available in conda-forge and prefix.dev).

  • Removed the deprecated lenskit.training.IterativeTraining base class in favor of lenskit.training.UsesTrainer.

  • Component class members have been renamed to no longer use the Scikit-Learn pattern of trailing _ in their names.

  • Removed the lenskit.stats.argtopn function (🐞 833). lenskit.data.ItemList.top_n() now uses a Rust-accelerated top-N implementation (⛙ 1049).

  • lenskit.data.ItemList.to_arrow() no longer accepts mappings for its optional columns argument, sequences of names are used instead (⛙ 1052).

New Features#

  • Added environment variables to TrainingOptions, along with the env_var() method to query them. This allows for open-ended configuration of training processes.

  • Added support for free-threaded Python, including binary distributions for Python 3.14t on Linux and macOS (🐞 916, ⛙ 1022).

  • Added support for parallel batch inference using thread pools on free-threaded Python (🐞 921, ⛙ 1025).

  • Added LightGCN support to FlexMF (🐞 1019, ⛙ 1033).

Performance Changes#

  • lenskit.data.RelationshipSet.co_occurrances() is much faster and uses parallel computation, at the expense of increased memory use in the symmetric (non-ordered) case (🐞 970, ⛙ 1007).

Minor Changes#

  • Pipeline type-checking for ArrayLike component inputs no longer works, due to a breaking change in NumPy 2.4. No LensKit components used ArrayLike as an input or output data type.

  • Pipeline component inputs with default values can now have missing inputs (🐞 1000, ⛙ 1001).

  • Rust progress updates now use a background thread to simplify logic and keep locks out of the work path (⛙ 1008).

  • Fixed a bug saving item list collections of emtpy item lists to Parquet files (🐞 1051, ⛙ 1052).