skrecsys¶

Recommender systems in the scikit-learn style, built on NumPy, SciPy and scikit-learn.

Every model has the same fit, partial_fit, recommend and predict interface, and works with scikit-learn's model selection tools. Around the models, skrecsys provides top-k ranking and beyond-accuracy metrics, a cross-validation splitter that keeps every test user and item in training, dataset loaders, and approximate nearest-neighbour indexes for large catalogs. The models are classical collaborative filtering (item kNN, BM25, RP3beta, EASE, SLIM, ALS, BPR) plus graph-based and sequential neural recommenders (SimpleX, XSimGCL, HSTU, Mamba4Rec).

The estimators are Python; the inner loops are not. The similarity kernels, the elastic-net and least-squares solvers, the BPR sampler, identifier encoding, sparse matrix assembly and top-k selection are all written in Rust and exposed through PyO3 as skrecsys._core. They parallelize with rayon and release the GIL while they run, and every recommender in skrecsys.recommendation scores and ranks a batch of queries inside one kernel, without ever materializing a dense score matrix in NumPy. The hot loops are written for the auto-vectorizer and use fused multiply-add where the hardware has it — NEON on Apple silicon and other aarch64, AVX2 and FMA on x86-64 through runtime dispatch. Nothing in the crate uses unsafe, and both crates #![forbid(unsafe_code)] so it stays that way.

The installed library is small: four runtime dependencies, all of them ones a scikit-learn user already has — NumPy, SciPy, scikit-learn and joblib, plus typing-extensions on Python 3.11 only. pandas is optional, and only needed for the dataset loaders' as_frame=True.