// the take
Surprise is a scikit-learn-style Python library for building and evaluating recommender systems based on explicit rating data (think star ratings, not clicks). It's aimed at people who want to prototype or benchmark classic algorithms like SVD, SVD++, KNN variants, and co-clustering without writing the cross-validation and evaluation plumbing themselves.
The API mirrors scikit-learn closely (fit/predict, cross_validate, GridSearchCV-style tuning), so anyone who's used sklearn is productive in minutes. The matrix factorization and similarity computations are written in Cython, not pure Python, which is why SVD on 100k ratings finishes in 6 seconds instead of minutes. Documentation is unusually thorough for a project this size — it spells out the actual math behind each algorithm rather than just listing method signatures. The built-in CV iterators and benchmark harness make it easy to honestly compare algorithms on the same folds.
It only handles explicit ratings — no implicit feedback (clicks, views, purchases) and no content-based filtering, which rules it out for most real-world recommender systems since those are built on implicit signals. The author froze feature development back in 2019 and has said so directly in the README: bugfixes only, no new algorithms, no deep learning methods. It's single-process and in-memory with no GPU support, so it doesn't scale past what fits on one machine — SVD++ on the 1M Movielens dataset already takes over 30 minutes per fold in their own benchmarks. Good for teaching, prototyping, or academic comparison of classic algorithms; not something to build a production recommender on.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →