recommenders-team/recommenders

Recommenders is an open-source collection of "best practices on recommendation systems", software that decides which products, articles, songs, or other items to suggest to each user. It is hosted by the Linux Foundation of AI and Data. Building a real recommender involves plumbing that does not change much between projects, plus a confusing landscape of algorithms. This repository packages both: a Python library called recommenders that handles the repetitive work, alongside runnable Jupyter notebooks that walk through each algorithm end-to-end. The notebooks are organised around five tasks that mirror a typical project: preparing and loading data in the shape each algorithm expects, building models, evaluating them with offline metrics, tuning hyperparameters, and operationalising a model in production on Azure. Inside the library, utilities handle dataset loading, train/test splitting, and metric calculation, and reference implementations are included for many algorithms, from classical matrix-factorisation approaches such as Alternating Least Squares on PySpark, to sequential convolutional models, content-based models that draw on a knowledge graph, and deep factorisation methods. Each algorithm typically comes with a "quick start" notebook and often a "deep dive" notebook explaining the mathematics. You would use Recommenders as a researcher comparing approaches on a benchmark dataset, as a developer prototyping a "people who liked this also liked…" feature, or as a team preparing to move a recommender from a notebook into production. The core package installs from PyPI, with optional extras for GPU training, Spark, development tooling, and experimental models.